Impact Magazine Spring 2018 by Impact Magazine from The Operational Research Society

D R I V I N G I M P R O V E M E N T W I T H O P E R AT I O N A L R E S E A R C H A N D D E C I S I O N A N A LY T I C S

SPRING 2018

MODELLING TRAFFIC FLOWS AT THE PORT OF DOVER Simulation helps keep Britain moving

O.R. HELPS KIDNEY PATIENTS Â© MikhailBerkut/Shutterstock.com

Algorithms help increase the number of transplants

PROVIDING PROTECTION AGAINST HAZARDOUS MATERIALS Queueing models inform decisions about vital protection for NHS staff

JOURNAL OF BUSINESS ANALYTICS The mission of the journal is to serve the rapidly growing and emergent community of business analytics both in academics and in industry/ with practitioners. We seek research papers that clearly address a business problem, develop innovative methods/ methodologies and use real-world data to show the how the problem can be solved.

Editors-in-Chief: Dursun Delen, Oklahoma State University, USA dursun.delen@okstate.edu Sudah Ram, Eller College of Management, USA sram@email.arizona.edu

T&F STEM @tandfSTEM

@tandfengineering

Explore more todayâ&#x20AC;Ś bit.ly/2g4s9YM

E D I TO R I A L Welcome to the first issue of Impact to be published by the OR Society’s new publishers, Taylor & Francis. I very much appreciate the work that has been put in to get this issue published on time. My special thanks are due to Katie Johnson and Ian Challand. We also welcome a new columnist, Louise Maynard-Atem, and look forward to her continuing contributions. She replaces my friend and colleague Mike Pidd, who has given us the benefits of his insights over the first six issues of this magazine, for which I have been very grateful. As this issue was in the final stages of assembly, TV screens were full of people in Hazmat suits in Salisbury, as they dealt with the aftermath of the poisoning of Sergei Skripal and his daughter, Yulia. How relevant, then, is the work highlighted in the article by Martin Utley and Luca Grieco, who developed analytical models to help those concerned with purchasing such suits? That article is only one of several on the health theme. David Manlove describes his important work, which has led to the increase in the number of kidney transplants. Tom Boness and Hannah Mayes explain how they have helped to improve the transportation of patients in Northern New South Wales, and Martin Pitt and Ken Stein tell us about their analytics group in the South-West of England, which is working with the NHS to make simulation and modelling techniques a key part of the decision-making process. Our lead article concerns work for the Port of Dover. It contains, I think, the only mention of Brexit (you can’t avoid it, can you?) as the modelling described in the article has been used, inter alia, to try to understand possible impacts of increased processing times on the Port’s performance. I trust you will enjoy reading this issue. Electronic copies of all issues continue to be available at https://www.theorsociety.com/Pages/Impact/Registration.aspx and are also at https://www.tandfonline.com/timp. For future issues of this free magazine, please subscribe at http://www.getimpactmagazine.co.uk/. Graham Rand

The OR Society is the trading name of the Operational Research Society, which is a registered charity and a company limited by guarantee.

Seymour House, 12 Edward Street, Birmingham, B1 2RX, UK Tel:+44 (0)121 233 9300, Fax:+44 (0)121 233 0321 Email: email@theorsociety.com Secretary and General Manager: Gavin Blackett President: John Hopes Editor: Graham Rand g.rand@lancaster.ac.uk

Print ISSN: 2058-802X Online ISSN: 2058-8038 www.tandfonline.com/timp Published by Taylor & Francis, an Informa business All Taylor and Francis Group journals are printed on paper from renewable sources by accredited partners.

OPERATIONAL RESEARCH AND DECISION ANALYTICS

Operational Research (O.R.) is the discipline of applying appropriate analytical methods to help those who run organisations make better decisions. It’s a ‘real world’ discipline with a focus on improving the complex systems and processes that underpin everyone’s daily life – O.R. is an improvement science. For over 70 years, O.R. has focussed on supporting decision-making in a wide range of organisations. It is a major contributor to the development of decision analytics, which has come to prominence because of the availability of big data. Work under the O.R. label continues, though some prefer names such as business analysis, decision analysis, analytics or management science. Whatever the name, O.R. analysts seek to work in partnership with managers and decision makers to achieve desirable outcomes that are informed and evidence-based. As the world has become more complex, problems tougher to solve using gut-feel alone, and computers become increasingly powerful, O.R. continues to develop new techniques to guide decision-making. The methods used are typically quantitative, tempered with problem structuring methods to resolve problems that have multiple stakeholders and conflicting objectives. Impact aims to encourage further use of O.R. by demonstrating the value of these techniques in every kind of organisation – large and small, private and public, for-profit and not-for-profit. To find out more about how decision analytics could help 1 IMPACT | 2018 your organisation make more informed decisions see www.scienceofbetter.co.uk. O.R. is the ‘science of better’.

CO N T E N T S 7

TRAFFIC MODELLING AT THE PORT OF DOVER Cliff Preston, Phillip Horne, Jesse O’Hanley and Maria Paola Scaparra describe their work to determine how the Port can best handle future volumes of traffic

HOW OPERATIONAL RESEARCH HELPS KIDNEY PATIENTS IN THE UK David Manlove reports how algorithms developed by University of Glasgow researchers led to an average annual increase of more than 20 kidney transplants and £5m savings for the NHS

PENCHORD: HELPING THE NHS MAKE INFORMED DECISIONS USING OPERATIONAL RESEARCH Martin Pitt and Ken Stein describe the work of PenCHORD with the NHS to make simulation and modelling techniques a key part of the decision-making process

OPERATIONAL RESEARCH IN SUPPORT OF CHARITYWORKS Howard Turner, Andrew Reeves and Tara Rowe describe how Pro Bono O.R. helped a charity achieve greater certainty, resource savings and improved processes

IMPROVING ENGLAND’S RESILIENCE TO INCIDENTS INVOLVING RELEASE OF HAZARDOUS MATERIALS

4 Seen Elsewhere

Analytics making an impact 13 Innovate OR Die

Louise Maynard-Atem looks at how soft O.R. techniques can be used to drive innovation across sectors and disrupt the status quo 20 Universities making an impact

Brief reports of two postgraduate student projects 37 Measuring efficiency through

Data Envelopment Analysis Emmanuel Thanassoulis and Maria Conceição A. Silva give insight into Data Envelopment Analysis: a method of choice for measuring and managing performance 46 Malgorithms

Geoff Royston argues that mathematical models should be our tools but not our masters

Luca Grieco and Martin Utley explain how they were able to use queueing models to inform decisions, valued at millions of pounds, concerning the purchase of protective suits

IMPROVING PATIENT TRANSPORT IN NEW SOUTH WALES Tom Boness and Hannah Mayes explain how their work allowed a more flexible transport service to be developed, with significant cost savings and benefits to patients

DISCLAIMER The Operational Research Society and our publisher Informa UK Limited, trading as Taylor & Francis Group, make every effort to ensure the accuracy of all the information (the “Content”) contained in our publications. However, the Operational Research Society and our publisher Informa UK Limited, trading as Taylor & Francis Group, our agents, and our licensors make no representations or warranties whatsoever as to the accuracy, completeness, or suitability for any purpose of the Content. Any opinions and views expressed in this publication are the opinions and views of the authors, and are not the views of or endorsed by the Operational Research Society or our publisher Informa UK Limited, trading as Taylor & Francis Group. The accuracy of the Content should not be relied upon and should be independently verified with primary sources of information. The Operational Research Society and our publisher Informa UK Limited, trading as Taylor & Francis Group, shall not be liable for any losses, actions, claims, proceedings, demands, costs, expenses, damages, and other liabilities whatsoever or howsoever caused arising directly or indirectly in connection with, in relation to or arising out of the use of the Content. Terms & Conditions of access and use can be found at http://www.tandfonline.com/page/terms-and-conditions

Reusing articles in this magazine All content is published under a Creative Commons Attribution-NonCommercial-NoDerivatives License which permits noncommercial re-use, distribution, and reproduction in any medium, provided the original work is properly cited, and is not altered, transformed, or built upon in any way.

TIMP_Prelim.indd 1

03-05-2018 17:01:57

SEEN ELSEWHERE

GETTING AN EDGE ON BORDER ACTIVITY

Apparently, the next incursion of terrorists into Indian soil is likely to happen in the early hours of the morning, around 2am, about eleven or twelve days after a herd of cattle meanders close to the border. How do Indian forces patrolling the border know this? A huge database of border movement information, including thermal imaging data, is combined with social media posts, hashtags, etc. Machine language algorithms and data mining techniques then help predict protests, riots and border incursions. ‘We have more than 20 terabytes of data on border movement which earlier used to be recorded in physical logbooks of soldiers. These include thermal images, instances of people going near the fence from across the border, activity at late night, etc,’ says Tushar Chhabra, cofounder of Gurgaon-based driverless truck company CRON systems, which helps the army in predicting borderinfiltration patterns. The cattle? Terrorists on the other side of the border usually send a herd of cattle to check for mines that the armed forces have placed on the border. Analysis of past data has shown that typically an incursion follows 12 days later. (See http://bit.ly/2g08ci7).

American artists born between 1900 and 1920, and their relation to the dates of death of the artists’ friends and family members. In ‘Death, Bereavement and Creativity’, to be published in Management Science (see https://doi.org/10.1287/ mnsc.2017.2850) they argue that personal unhappiness, particularly that experienced in times of mourning or bereavement, can cause a decrease of about 35% in the value of an artist’s work. The authors also found that there was no statistically significant difference in terms of whether the death involved a parent, a sibling or a friend, and this decrease in the value of their work typically did not extend beyond that one-year time frame, and that these paintings were much less likely to be included in a museum collection.

TRUST IN NUMBERS

‘TORTURED ARTISTS’ AND VALUE OF THEIR WORK

WILL SHE REMAIN OR LEAVE?

Kathryn Graddy of Brandeis University and Carl Lieberman of Princeton University have studied the prices of more than 10,000 paintings produced by 33 French impressionist artists and more than 2,000 paintings by 15

ADP, a USA based provider of human resources management software and services, has developed an algorithm which can predict which staff are likely to leave and also suggest which staff are the best ones to retain by incentivising

IMPACT © THE OR SOCIETY

and promotion. It does this by running in the cloud and continuously monitoring the performance of individual employees. ‘We are very confident in our flight-risk predictor,’ said Stuart Sackman, CIO/CTO and corporate vice president, global product and technology, at ADP. Sackman advises that employers temper flight risk with high performance analysis to determine how best to address the issue with individual employees. If a prized employee is at risk of leaving, employers and managers might be better served to try to find ways to try to keep them on staff. However, if the employee is already under-performing, perhaps parting ways sooner is better. ‘The insights we provide are not just complex tools but help people make better decisions,’ said Sackman. ‘Based on real Big Data’ stored and analysed in the Cloud. See: http://fxn.ws/2cr7lJ7

Sir David Spiegelhalter entitled his Presidential Address to the Royal Statistical Society ‘Trust in Numbers’. He was concerned about claims both of a reproducibility crisis in scientific publication and of a post-truth society abounding in fake news and alternative facts. By considering the ‘pipelines’ through which scientific and political evidence is propagated, he considered possible ways of improving both the trustworthiness of the statistical evidence being communicated and the ability of audiences to assess the quality and reliability of what they are being told. Trust in Numbers is published in Journal of the Royal Statistical Society: Series A (Statistics in Society) (2017), 180, 948–965.

COUNTERING HUMAN TRAFFICKING

Probably more than 30 million people worldwide are trapped in some form of modern day slavery, whether as sex workers, beggars, child soldiers, domestic workers, factory workers and labourers across a very wide range of industries. Human trafficking is essentially a supply chain. The human traffic, just like any other ‘commodity’, moves through a network designed to meet demand. This commodity is cheap, vulnerable and illegal labour. During their trafficking operations, traffickers leave a data trail, but such trails are rarely overt, often they are faint or broken, yet still visible despite traffickers’ efforts to operate unseen and off the grid. Agencies do their best with limited resources to break the supply chain and free its traffic, but funding is often a scarce resource for such work. Anti-trafficking organisations with limited resources may therefore benefit from the insight data analysis can provide. Fortunately, algorithmic extraction of useful data is now possible and can be used to assist law enforcement agencies. ‘Bits’ of information extracted from these trafficking tracks can be used to uncover the mechanisms used in the distribution of victims, traffickers, buyers and other exploiters, and disrupt the supply chain wherever and whenever possible. A team of researchers in the USA provide insight into the application of data science technologies to help thwart human trafficking. They identify opportunities for O.R. and analytics in the war against human trafficking, and demonstrate that they can genuinely make a difference! (See European Journal of Operational Research (2017), 259, 733-745).

GENERATING GUITAR SOLOS USING INTEGER PROGRAMMING

Researchers Nailson dos Santos Cunha, Anand Subramanian and Dorien Herremans have presented a framework, in the Journal of the Operational Research Society, for computer-aided composition that uses exact combinatorial optimisation methods to generate guitar solos from a newly proposed data-set of licks over an accompaniment based on the 12-bar blues chord progression. An integer programming formulation, which can be solved to optimality by a branch-and-cut algorithm, was developed for this problem whose objective is to determine an optimal sequence of a set of licks. Apparently, an empirical experiment with 173 participants show that the solos whose licks were optimally sequenced were significantly more enjoyed than those randomly sequenced. See http:// tandfonline.com/doi/full/10.1080/0160 5682.2017.1390528.

DEALING WITH BIG DATA

Big data has caught the attention of business leaders, though there is still widespread confusion as to how to treat such data. Are we collecting the right data? Do we have the capacity to store the data we collect? Is the data being analysed correctly? Researchers in the USA have written about lessons they learned from a company dealing with big data. (See https://doi.org/10.1287/ inte.2017.0890). They learnt that: (1)

multinational companies can face problems because of a basic lack of analytical expertise; (2) although many companies possess copious amounts of data, they lack the ability to process these data in any meaningful manner; (3) analytical methods need to be developed incrementally, sometimes beginning with fundamental concepts and tools; (4) although starting with small samples is a common way to ensure proof of concept, processes generally need to be scalable, and users need to be aware of their limitations; and (5) the trend toward employing an in-house computational group will have increasingly important ramifications for the company’s overall operational and financial health. MATHEMATICAL TREATMENT OF CANCER

Dr. Alexander Anderson, chair of integrated mathematical oncology at the Moffitt Cancer Center, describes the dynamics of cancer mathematically in the form of non-linear equations. ‘Understanding the dynamics of a cancer, how it grows and evolves is not something that’s intuitive; it’s something that you can write down in equations and those equations can describe this non-linear and nonintuitive behaviour better than a human could,’ says Dr. Anderson. Different treatments can be run through these algorithms to determine the one best suited to the given cancer – patient combination. Rather than zapping the sensitive cells, which tends to leave behind the resistant ones, his proposed treatment is to apply ‘just enough’ repetitively to keep the tumours under control – to treat cancer like a chronic disease. Most cancer treatments aim to cure cancer, but Anderson argues that this might not always be the best course of treatment.

IMPACT | SPRING 2018

THE NEW WORLD OF DATA

Writing on the power of data in Lancaster University Management School’s Fifty Four Degrees, alumnus Craig Boundy, Chief Executive Officer of Experian North America says that ‘businesses are using big data for enhancing customer experiences, streamlining existing processes, creating more targeted marketing and reducing costs. Rather than being the preserve of Chief Information Officers, big data

IMPACT | SPRING 2018

projects are increasingly being driven by business unit heads as awareness grows of the potential for more data-driven decision-making’. He continues: ‘Businesses need to be aware of pitfalls. It is the ‘dark side’ of data – an inability to effectively manipulate and interpret the data, act on it and prove ROI – that is holding organisations back from making full use of its potential. Ultimately, there are greater risks to organisations from not being part of the new world of data than from the risks of trialling an approach’.

– providing distractions is likely to prove counter-productive. For instance, the practice, commonly employed at restaurants, of providing customers with a menu to view while they wait achieve their intended effect of reducing dissatisfaction due to long waits, but they also result in lower consumption, suggesting that ‘higher customer satisfaction’ is not always the same as ‘higher revenue.’ So next time you are in a long line, if you do not want to make a large purchase that you might later regret, it is suggested that you distract your attention as much as you can. (See https://doi.org/10.2139/ ssrn.3007786). AI AND DECISION-MAKING

GET IN LINE

In a series of experiments conducted in the lab and in the field, researchers from Georgetown, USA, found strong evidence to indicate that the longer we are kept waiting in a queue at a shop or restaurant the more we will buy. For example, diners would want their dinner service to last longer if they had waited longer for a table; shoppers would buy more T-shirts after having a longer wait at a clothing store during sales; and people would play more rounds of an arcade game if they had to wait longer for earlier customers to finish. They suggest this is driven by mental accounting: a larger purchase ‘justifies’ the fixed cost of the long wait. However, this apparently only works if the wait is perceived to be long

Ray Dalio, founder of Bridgewater Associates, the world’s largest hedge fund, whose new book ‘Principles’ is a bestseller, said in Forbes magazine’s 100th Anniversary Issue under the heading, ‘Lessons and Ideas by the 100 Greatest Living Business Minds’: ‘I think the most important issue that will reshape our lives in the years ahead will be how man-made and artificial intelligence compete and work together. My views have been colored by experiences with algorithmic decisionmaking over the last 30 years, which have been fabulous. But it’s a two-edged sword. I have learned that by thinking through my criteria for making decisions, writing them down as principles and then expressing them as algorithms so that the computer thinks in parallel with me, I can make much better decisions than I could make alone. It has helped us to have an idea meritocracy that produces collective decision-making. But our path to doing this was to work with the computer to gain deep understanding.’

‘When we give continuous, dose-dense therapies, what you’re treating is the sensitive cells and you’re treating so long that you wipe out all the sensitive cells and all that you leave behind are the resistant cells,’ explains Anderson. When there are only resistant cells left, treatment no longer works. Instead, the goal of Anderson’s treatment isn’t to cure but rather to stave off metastasis and prolong life. The mathematical models let doctors run different treatment scenarios through an algorithm to determine the best one, without ever actually having to test anything on the patient themselves. Using the algorithms, Anderson is able to find the sweet spot, not too much treatment and not too little, so that the treatment can be applied over and over again and ultimately control the tumour. More research is needed before this approach makes its way into general practice, but Anderson hopes that eventually it will be used in hospitals and cancer centres around the world. ‘Tailoring mathematical models to each individual patient’s cancer and using that as a predictive tool shouldn’t be that far out. We’re talking five years probably,’ he says. (See http://bit. ly/2xIHH8B).

T R A F F I C M O D E L L I N G AT T H E P O R T O F D OV E R CLIFF PRESTON, PHILLIP HORNE, JESSE O’HANLEY AND MARIA PAOLA SCAPARRA

THE PORT OF DOVER has undergone many reincarnations over the centuries: from a fortified port complete with lighthouse in the first century AD, to a military Cinque Port in the middle ages, to the ferry and hovercraft terminal of the late twentieth century. Dover’s principal role now is as a Roll-on, Roll-off (Ro-Ro) Ferry Terminal, in which 2 ferry companies (P&O and DFDS) between them make up to 60 round trips a day to the French Ports of Calais and Dunkerque. They

carry over 2.6 million lorries, 2 million cars, and 12 million people a year. The economic value in goods handled through the Port is up to 17% of the UK’s overall trade in goods. Based on 2016 projections, freight traffic is expected to increase by up to 40% in the next 30 years. However, the Dover Eastern Docks Ferry Terminal is small, around half a square kilometre, and expansion is challenging since it is hemmed in by the sea, the White Cliffs of Dover, and Dover town.

more effectively. Fluidity of movement and utilising the space we have as efficiently as possible is a fundamental aspect of master planning that benefits both our customers and the community. Handling 17 per cent of the UK’s total trade in goods, this work has helped us keep Dover, Kent, and Britain moving for both businesses and consumers, as well as plan for the future.’ – Rikard Bergstrom, General Manager of Engineering, Port of Dover.

MODELLING TRAFFIC FLOWS

In 2015, the Port and University of Kent Business School began a 2-year Knowledge Transfer Partnership (KTP) funded by Innovate UK. One objective of the KTP was to determine how the Port could best handle future volumes of traffic based on modelling current and future traffic flows. The aim of modelling was to identify potential bottlenecks in the system and evaluate a range of options and investments for the future.

One objective of the KTP was to determine how the Port could best handle future volumes of traffic based on modelling current and future traffic flows

‘The Knowledge Transfer Partnership between the University of Kent and Port of Dover has given us a better understanding of the dynamics of traffic flow and how to handle these flows

FIGURE 1 SCHEMATIC OF THE PORT OF DOVER EASTERN DOCKS (OUTBOUND TRAFFIC)

IMPACT | spring 2018

this work has helped us keep Dover, Kent, and Britain moving for both businesses and consumers, as well as plan for the future

Early in the KTP, a decision was made to use discrete-event simulation to model road traffic flows at the Port. The discrete-event approach was preferable to agent-based modelling, since the Port consists of a series of process steps, each preceded by orderly queues rather than a melee of interacting vehicles. The layout and processes of the Port are shown in Figure 1. Traffic heading for the continent approaches the Eastern Docks from the M20/A20 and M2/A2 roads. On entering the Port, the first checkpoint is the French border control (Police aux Frontières, or PAF), followed immediately by Kent Police. Next, there is ferry operator Check-In, after which vehicles are marshalled in an Assembly Area prior to embarkation onto ferries. The unusual arrangement, in which French immigration entry checks into the Schengen Area are performed in the UK, has been in place since the Le Touquet Treaty of 2003. As can be seen in Figure 1, there are 3 areas in the Port where traffic queues are formed: the marshalling/assembly area

next to the ferry berths, a small plaza in front of the check-in booths, and a large ‘Buffer Zone’ (see Figure 2) – lanes at the entrance to the Port capable of holding 4km of traffic that was created in 2016 as part of an £85m project co-funded by the European Union. From an applied research perspective, the Port has interesting dynamics, consisting of a series of processes and associated queues in series. The arrival rate of vehicles at the Port is periodic, rather than constant, and processing rates are fairly efficiently matched with average arrivals such that queues frequently ebb and flow as the arrival rate exceeds then falls back below processing rates. There are two further queue management systems that the Port can make use of. The first is called the Dover Traffic Access Protocol (TAP), in which freight vehicles (and only freight) can be held on the A20 approach with traffic lights. Dover TAP ensures the free-flow of tourist and local traffic along the A20 and through the town by controlling freight when demand is at its highest. TAP has been in place since 2015, first as an experiment, but following its initial

success is now a permanent option operated by the Port with the consent of Highways England. The final queue management control is the well-known ‘Operation Stack’, an option of last resort in which the M20 motorway is used as a lorry queueing area. Operation Stack is only triggered in exceptional circumstances – not at all during 2016/17 and last used in 2015 due to unprecedented events, including the Port of Calais blockade and closure and migrant disruption of the Eurotunnel train line. Operation Stack is estimated to cost the UK up to £2m per day in direct costs and lost productivity.

MODELLING THE PRESENT DAY

The first step in the KTP was to develop a Simul8 model (Figure 3) of presentday operations at the Port for validation. The Port is data-rich, with excellent historical data on ferry schedules, vehicle/passenger carryings, and arrival rates for different traffic types. This was supplemented by direct timings of processes where needed. The Port has

also invested in Blip Systems’ BlipTrack™ system, which anonymously senses mobile phone and satellite navigation signals that can be used to monitor transit and dwell times of traffic on the approach to and through the Port. The accuracy of the baseline model was checked under normal conditions as well as under a variety of extremes (check-in IT failure, ferry re-fit schedule causing low daily capacity on routes, and heightened security procedures). The model was considered adequate for its intended use. Modelling the Port’s present-day operations provided a number of useful insights, some of which were not immediately obvious. The first was that the physical infrastructure of the Port system and controls could handle present-day traffic volumes quite well, so that when, for example, check-in processes caused excess queues, it was generally due to technical issues rather than physical capacity limits (e.g. IT problems requiring slow manual check-in processing or staff shortages due to illness). Second, the modelling showed how incredibly useful the Buffer

FIGURE 2 THE PORT OF DOVER BUFFER ZONE WITH A MIX OF TRAFFIC TYPES.

IMPACT | spring 2018

FIGURE 3 BASELINE SIMUL8 MODEL OF THE PORT OF DOVER EASTERN DOCKS

Zone is, since it can provide a selective buffer for different traffic types and ferry operators, as seen in Figure 3, where freight and tourist traffic for each ferry operator are handled separately. Additionally, being at the entrance to the Port, the Buffer Zone provides resilience against problems at any point in the system, in contrast to say the Assembly Area, which cannot mitigate against problems at border controls or check-ins.

MODELLING FUTURE GROWTH

The next step in the modelling study was to consider future traffic volumes. Three forecasts: expected, optimistic, and pessimistic were constructed to bracket possible future growth, based on varying assumptions and the use of sensitivity analysis to pinpoint future pinch-points. Part of the work also considered how best to use two parcels of land which could be converted (at a cost) from other purposes to operational space. A key performance measure to assess different options was ‘TAPs per year’ – the number of times the capacity of the Port itself is exceeded, resulting

IMPACT | spring 2018

in the need to trigger Dover TAP. This is not primarily a financial metric, but rather the degree of inconvenience to the Port, its customers, and the local community. Modelling showed a mixture of obvious results and some surprises. In the former category was the observation about the importance of daily uplift – the total amount of vehicles ferries can carry per day. This must ultimately match or exceed daily traffic arrivals to avoid queues, although, as mentioned previously, it is quite feasible for arrivals to exceed uplift transiently within a day, causing occasional traffic queues. This is unsurprising since the ferries provide the only ‘sink’ which removes outbound traffic from the system. More surprising was that even allowing for considerable increases in traffic volumes, the physical infrastructure of the Port could suitably cope with future demand, with the proviso that staffing would need to increase proportionally to handle increases in traffic. However, as volumes increase, resilience falls. For example, an interruption to check-in would cause

queues to form much more rapidly under future traffic volumes than in the present day. Essentially, any dynamic process would be sped-up. For the land which could be converted, the best possible use would be flexible space, ideally with the same characteristics of the Buffer Zone. Modelling identified a costed, sequenced set of investments that optimally addressed traffic growth, together with critical traffic volume thresholds that should trigger investments in the future. The overview being given may make it sound as if the KTP was a routine exercise in simulation modelling. This, however, would very much underplay how work had to quickly adapt to real-world events as and when they occurred. The first such instance was in November 2015, one week into the KTP project. On the night of the 13th, terrorists attacked numerous targets across Paris, including the Bataclan Theatre and Stade de France Stadium. The resulting change in security procedures had a major knock-on effect on the Port, resulting in PAF processing

becoming a key limiter of vehicle throughput. Over time, the Port has adapted to this change and continues to use modelling to investigate adjustments to security levels on traffic flows. The second key event to influence the modelling study has yet to play out fully. Early in the KTP, an influence diagram was built representing both permanent and transient factors affecting traffic flows. One factor, considered unlikely, was ‘that Britain should vote to leave the EU’. As we now know, on 23 June 2016 that risk crystallized. As of early 2018, what ports and airports need to do to fulfil new requirements imposed by Brexit are still unknown. However, the Port has used modelling to try to understand possible impacts of increased processing times on Port performance, both inbound and outbound. This has shown, for example, that as little as 2 minutes per freight vehicle added on to check-in times would result in massive queues and, since the Port has extremely limited space available, queues would be external to the Port, thus necessitating increased use of TAP or, worse still, Operation Stack.

IMPACTS OF THE WORK

Overall, the KTP project has had several positive impacts for the Port, both tangible, such as saving money and resources, and intangible, such as facilitating dialogue with government agencies.

the project has had several positive impacts for the Port, both tangible, such as saving money and resources, and intangible, such as facilitating dialogue with government agencies

But arguably the most lasting of benefits is the development of a few simple principles that apply under virtually any future circumstances, as detailed below.

the most lasting of benefits is the development of a few simple principles that apply under virtually any future

The value of flexibility. Given the intrinsic variability in traffic flows (transient over a day, week, or season) no one fixed configuration of the Port’s hard infrastructure could possibly be optimal all the time. Any investments which increase flexibility, such as using temporary or relocatable structures, are to be welcome. Conversely, investments which reduce flexibility run the risk of being regrettable at some point in the future. Ask key questions when evaluating any future plans. Does it increase capacity? Does it improve fluidity? Does it enhance flexibility? Does it increase resilience? Recognising the difference between the ‘anatomy’ of the Port (i.e. physical infrastructure) and its ‘physiology’ (i.e. how it works). More specifically, it is preferable, where feasible, to adapt performance by changing behaviours through incentives, nudges, and other means, rather than resorting to changes in physical infrastructure. Understanding the importance of modelling to support planning. The Port is a dynamic system influenced by many external factors (e.g. road conditions, weather, economics, and socio-political events) and relies on highly-skilled and experienced staff to manage the system. Forecasting and analytics serves an important role in supporting real-time,

evidence-based decision-making of Port staff. Cliff Preston received an MSc in Management Science from the University of Kent in 2014 and is now working part-time as a Strategy Analyst at the Port of Dover. He is also a freelance decision analyst working mainly in the pharmaceutical and biotech industries. Phillip Horne is a graduate in Mathematics from the University of Kent. He heads the Business Optimisation Team at the Port of Dover and is responsible for coordinating the Port’s 30-year Master Planning process. Jesse O’Hanley is a Reader in Environmental Systems Management in the Kent Business School, University of Kent and a member of the Centre for Logistics and Heuristic Optimisation. He is the winner of the 2015 EURO Excellence in Practice Award for work on optimising river connectivity restoration. Maria Paola Scaparra is the head of the Management Science group at Kent Business School and a member of the Centre for Logistics and Heuristic Optimisation. Funding for the KTP was provided in part by Innovate UK. This support is gratefully acknowledged. The authors would also like to thank Port of Dover CEO Tim Waggott, KTP Facilitator Rikard Bergstrom, KTP Advisor Terry Corner, former General Manager of Strategy & Risk Management Timothy Godden, and the University of Kent Innovation and Enterprise Team for all their contributions, as well as colleagues at the Port of Dover who helped provide data, advice, and assistance throughout the project.

IMPACT | spring 2018

I N N OVAT E O R D I E Louise Maynard-Atem

WE ARE CURRENTLY LIVING in a very interesting period of history; the ever-increasing rate of technology change and the sheer volume of data (as well as our ability to process and glean insights from it) is driving massive disruption across all industries. Longstanding companies are now having to redefine their value proposition and relationship with their customers, in order to keep pace with new entrants to the market as well as incumbent competitors. I’m sure you’re all familiar with the tale of entertainment rental service Blockbuster, and how their inability to move with the prevailing trend of streaming video content led to their eventual demise. All the while Netflix, who started out as a postal DVD service, were able and willing to pivot towards a technology on the rise and go on to dominate the market. I currently work in the innovation team of a large corporate organisation, and I hear some form of the Blockbuster vs Netflix story at least once a day. This, and other cautionary tales like it, emphasise the need for constant re-evaluation of the value that organisations provide to their customers and how they can better deliver this value. On the surface of things, my current role has a lot more to do with product design and management than it does with O.R. The purpose of the team that I work in is to develop new products and services that will ensure the medium and long-term success of the organisation. To do this we need to understand the problems that these products and services will solve, and that’s where my O.R. background really starts to add value. There are wealth of soft O.R. approaches and techniques that can support product innovation process; I have seen a number of these used, to great success, by a wide range of organisations and have already incorporated a number of them into my own ways of working.

THE INNOVATION PROCESS

Innovation and process are perhaps two words that don’t sit naturally together, with one conjuring images of free-thinking and new ideas, whilst the other lends itself to structure and certainty. The purpose of my role, and other innovation teams that are increasingly common in large organisations, is to take new ideas that are not yet fully formed and turn them into the business-as-usual of tomorrow. Another key point is to make this repeatable – thus requiring a supporting process. There are a number of different versions of what the innovation process looks like, however they all incorporate a set of common themes as shown in Figure 1: • Identifying the problem – the most important stage of the process is to really understand what problem you are trying to solve, and then reframing that problem to ensure you can generate as many ideas as possible in the next stage. • Ideation – once you have identified the problem you’re trying to solve, the next phase is to come up with as many ideas on how best the problem can be solved. The ideas will then need to be prioritised accordingly. • Research & Validation – once you have decided on the top idea(s) that you want to pursue, you need to test this with the market to ensure that you’re solving a problem that is important to customers. It’s also important to assess the technical feasibility of the idea at this stage. • Prototyping/MVP – only at this point do you start to build any functionality and it is vital that this is done in conjunction with customers/users. This stage helps to assess the technical feasibility of the idea as well as what the user experience should be. • Launch – once the idea has been validated from both a technical and market perspective, it is time to launch the product at scale and work towards driving adoption, both internally and externally.

The purpose of innovation teams that are increasingly common in large organisations, is to take new ideas that are not yet fully formed and turn them into the businessas-usual of tomorrow

It is also important to recognise the feedback element of the process; at every stage we are building confidence in the assumptions that we’ve made along the way, and feeding the

FIGURE 1 THE INNOVATION PROCESS

learnings that we achieve back into the process in order to minimise the risk and uncertainty.

PROBLEM STRUCTURING AND REFRAMING

The innovation space is often, incorrectly, considered to be solution or technology driven, though it is a firm understanding of the problem that you are trying to solve that is crucial to success – hence that is the first step in the innovation process shown. It is also important to understand the range of actors that are involved in the problem space, and take into account their different viewpoints. Soft systems methodology is an obvious choice for the initial problem structuring exercise that takes place at the very start of the innovation process. Soft Systems Methodology (SSM) traces its roots back to Lancaster University, when researchers were trying to apply systems engineering approaches to management/business problems. The methodology was born out of the fact that stakeholders have very different views of the ‘system’ and therefore define the problem in different ways. The seven stages of SSM, as shown in Figure 2, form the first stage of the innovation process and help us to achieve a common understanding of the problem that we can ideate on. In working to define the problem that we are solving, we need to understand the context and environment in which we

IMPACT | spring 2018

will solve it. Given we are aiming to develop solutions that will exist in a future that has not yet arrived, we need some structured way of thinking about what that future will look like; which is where scenario planning comes in. Scenario planning, which has its origins in military planning, is widely used in corporate innovation teams to generate multiple versions of the future. It allows us to consider factors that are very difficult to quantify including social/demographic, technological, economic, environmental, political, legal and ethical (STEEPLE) factors. When used in conjunction with systems thinking approaches (e.g. SSM), scenario planning can help us to develop plausible futures because the causal relationship between different factors can be demonstrated.

Soft systems methodology is an obvious choice for the initial problem structuring exercise that takes place at the very start of the innovation process

Defining the problem is long process, and it is valuable to try and reframe the problem a number of times before starting to think about solutions. To use a trivial example, if we ask ourselves ‘what is the sum of 5 + 5?’, there is

FIGURE 2 THE SEVEN STAGES OF SSM (ADAPTED FROM P. CHECKLAND, SYSTEMS THINKING, SYSTEM PRACTICE, 1981)

only one answer. However, if I were to ask ‘what two numbers add up to make 10?’, I open myself up to a whole range of solutions. Problem reframing is particularly valuable, as people often approach problems with their own biases. By reframing the issue, it encourages people to think beyond their own natural bias when trying to develop solutions.

Using Multi Criteria Decision Analysis allows us to quantitatively compare different ideas and drives more defensible decision-making behaviours

PRIORITISATION OF IDEAS

Once we have understood the problem, using a range of problem structuring methods, and developed a range of ideas that could provide solutions, we need a way to prioritise the ideas and decide which one should move towards implementation. Again, O.R. techniques are heavily used in this space to evaluate multiple criteria to aid decisionmaking. Multi Criteria Decision Analysis (MCDA) is particularly useful in identifying the interests of stakeholders/decision makers, which are often linked to the value of the opportunity and the strategic alignment with the wider portfolio. These interests are weighted in terms of importance and the different ideas are rated in line with the criteria. Using MCDA allows us to quantitatively compare different ideas and drives more defensible decision-making behaviours.

This element is particularly important in large organisations because any new innovation will have to exist in the context of existing solutions, and must therefore be aligned to the same vision as the wider offerings.

KEY TAKEAWAYS

To echo the point I made at the beginning of this article, the current pace of change is great because of the availability of data and the lower barriers to entry from a technology perspective. Large organisations are continuously having their market dominance challenged by an increasingly crowded competitive landscape, and need to be to move at speed in order to maintain their positions. The advent of corporate innovation teams, partnerships with start-ups and early stage investment (as well as more traditional merger and acquisition activities) are testament to how serious the threat of disruption is. The techniques listed above are just a few examples of how innovation professionals are using O.R. techniques to bring about changes to products and services in organisations of all sizes. The widespread adoption of these techniques (albeit sometimes under different names) speaks to massive applicability that O.R. has across a broad range of areas, and the drive towards evidence-based decision-making in business. Louise Maynard-Atem is a business consultant in the Futures and Innovation team at BAE Systems. She is an active member of the O.R. Society and serves as chair of the Early Careers Advisory Group. She is also an advocate for STEM activities, volunteering with the STEMettes and The Access Project.

IMPACT | spring 2018

DAVID MANLOVE

ALGORITHMS DEVELOPED BY UNIVERSITY OF GLASGOW RESEARCHERS led to over 200 more kidney transplants taking place between 2008 and 2017 than is estimated to have been the case had a previous algorithm continued to be used. This potentially saved the NHS around £52M over a 10-year period.

KIDNEY EXCHANGE

Kidney failure has a devastating impact on patients’ lives, and long-term survival

rates after transplantation demonstrate a doubled or tripled life expectancy compared to dialysis. NHS Blood and Transplant (NHSBT) estimates that over 37,500 people in the UK have end-stage renal failure; nearly 21,000 are on dialysis. As of 31 March 2017, there were 5233 patients on the transplant list for a donor kidney. The number of kidney transplants carried out each year is much less than this number: 3347 transplants took place between 1 April 2016 and 31 March 2017, of which 1009 were from living donors.

H OW O P E R AT I O N A L RESEARCH HELPS KIDNEY PAT I E N T S I N T H E U K

A patient may have a willing donor who is blood-type or tissue-type incompatible with them. In the past, typically that would have meant that the donor would have been unable to help their loved one. However, following the introduction of the Human Tissue Act in 2006, there is now the legal framework to allow transplants between strangers, thus opening up new possibilities for living donor transplants. For example, through a paired kidney exchange (PKE), a group of two or more kidney patients can swap their willing but incompatible donors with one another in a cyclic fashion, so that each patient can receive a compatible kidney.

In a number of countries, centralised programmes (also known as kidney exchange matching schemes) have been introduced to help optimise the search for PKEs. These countries include the UK, USA, the Netherlands, Australia, South Korea and many others around the world

Most commonly, PKEs involve either two or three patients, in which case they are called pairwise exchanges and 3-way exchanges respectively (see Figure 1 for illustrations of these types of PKEs). In a number of countries, centralised programmes (also known as kidney exchange matching schemes) have been introduced to help optimise the search for PKEs. These countries include the UK, USA, the Netherlands, Australia, South Korea and many others around the world. In general, it is logistically challenging to carry out the transplants involved in a PKE when the number of

FIGURE 1 HOW THE KIDNEY EXCHANGES WORK. IN EACH PKE, p REPRESENTS A PATIENT AND d REPRESENTS A DONOR. IN THE CASE OF THE PAIRWISE EXCHANGE, FOR EXAMPLE, d1 DONATES A KIDNEY TO p2 IN EXCHANGE FOR d2 DONATING A KIDNEY TO p1

pairs involved in a single such exchange is large. This is mainly because all operations have to be performed simultaneously due to the risk of a donor reneging on his/her commitment to donate a kidney after their loved one has received a kidney. But also, longer PKEs involve more participants, and therefore carry a higher risk that the whole cycle will break down if one of the donors or patients involved becomes ill. Mainly for these reasons, most centralised programmes only allow PKEs to involve pairwise and 3-way exchanges. Even 3-way exchanges, for example, require substantial coordination, involving six operating theatres and surgical teams scheduled on a single day (for three nephrectomies and three transplants).

COLLABORATION

In early 2007, NHSBT (formerly UK Transplant) set up the UK’s national kidney exchange matching scheme, now known as the UK Living Kidney Sharing Schemes (UKLKSS), to identify optimal sets of PKEs from among the patient and donor data on the NHSBT database. The algorithm that they developed was only capable of finding pairwise exchanges and could only handle datasets of up to 100 potential transplants.

In May 2007, the author and Dr Péter Biró (both of the University of Glasgow) developed a novel approach involving graph matching algorithms which enabled optimal sets of PKEs, involving pairwise and 3-way exchanges, to be identified. They also significantly increased the capacity of the algorithm to deal with larger datasets of up to 3000 potential transplants. Simulations that they ran using their implemented algorithms indicated the likely benefit, in terms of numbers of additional transplants, of allowing PKEs to involve 3-way exchanges in addition to pairwise exchanges. Following this initial research, NHSBT took the decision in April 2008 to allow PKEs to include 3-way exchanges. The introduction of these exchanges meant that finding an optimal set of PKEs had become a provably hard problem (known technically as an NP-hard problem), making it challenging to solve efficiently. This software was used at quarterly matching runs of the UKLKSS between July 2008 and October 2011 to find optimal sets of kidney exchanges. Between 2010 and 2011, an improved version of the Glasgow software was written in collaboration with Dr Gregg O’Malley (also of the

IMPACT | spring 2018

FIGURE 2 DOMINO PAIRED DONATION (DPD) CHAINS. IN EACH CHAIN, a1 IS AN ALTRUISTIC DONOR, WHO DONATES A KIDNEY TO p2, IN EXCHANGE FOR d2 DONATING A KIDNEY, ETC., WITH THE FINAL DONOR IN EACH CHAIN DONATING A KIDNEY TO A PATIENT ON THE DECEASED DONOR WAITING LIST (DDWL)

University of Glasgow), using integer programming, a technique often used in Operational Research. This also addressed some changes in the established criteria for matches, which are described in more detail below. Moreover, the software was extended to deal with domino paired donation (DPD) chains. Such chains are triggered by altruistic donors, who wish to donate a kidney but do not have a specific patient in mind, together with incompatible donor-patient pairs, with the final donor donating to the deceased donor waiting list (see Figure 2 for illustrations of DPD chains). DPD chains have featured in the matching scheme since January 2012. Drs Manlove and O’Malley developed an in-house version of the software for NHSBT, allowing them to conduct the searches themselves and speed up response times. This software was delivered to NHSBT in June 2011, and has been used to find optimal solutions for quarterly matching runs of the UKLKSS since January 2012.

involve maximising the overall number of transplants, mitigating the risk associated with 3-way exchanges and long chains (by minimising their use where possible), and maximising the overall score of the identified PKEs and chains. The score of a potential solution is calculated by a scoring function used by NHSBT, which takes into account factors such as waiting time (based on the number of previous matching runs that a participant has been unsuccessfully involved in), sensitisation (which roughly corresponds to how difficult to match a patient is), HLA mismatch levels between a donor and

patient (which relate to levels of tissue-type incompatibility) and points relating to the difference in ages between donors. (See the article by Manlove and O’Malley referenced at the end for a more detailed description of the optimisation problem and the algorithm used to solve it.) As indicated previously, the algorithm is based on integer programming. It is implemented in C++ and uses the COIN-CBC solver. Figure 3 gives an illustration of the optimisation problem, together with an optimal solution found by the algorithm for the July 2015 data-set. The diagram shows a representation of the problem as a directed graph, consisting of vertices (depicted by circles, representing donor-patient pairs and altruistic donors) and arcs (depicted by arrows, representing compatibilities between donors and patients). An optimal solution found by the algorithm is highlighted using solid arcs, comprising seven pairwise exchanges, five 3-way exchanges and six long chains. To date, the algorithm has

THE OPTIMISATION PROBLEM

The algorithm developed by the University of Glasgow researchers for the UKLKSS currently has to solve a complex optimisation problem involving five optimality criteria that are optimised in a hierarchical fashion. These criteria

IMPACT | Spring 2018

FIGURE 3 THE UNDERLYING DIRECTED GRAPH CORRESPONDING TO THE JULY 2015 DATASET, TOGETHER WITH AN OPTIMAL SOLUTION. TURQUOISE AND BLUE VERTICES ARE ALTRUISTIC DONORS (THE FORMER TRIGGER LONG CHAINS WHILST THE LATTER ARE UNMATCHED). TURQUOISE AND RED VERTICES BELONG TO PKES AND CHAINS, WHILST PINK VERTICES REPRESENT UNMATCHED DONORS AND PATIENTS

found optimal solutions for the datasets arising from all quarterly matching runs within 7 seconds. Since delivering the software containing the optimal matching algorithm to NHSBT, subsequent work has involved carrying out simulations to measure the effectiveness of the current optimality criteria, and has also led to improved integer programming formulations combined with column generation techniques to handle larger datasets and longer DPD chains. A key aim of this scalability work is to ensure that the matching algorithms are capable of anticipating future challenges that may emerge from larger pool sizes and more complex optimality criteria.

stated ‘Since July 2008, we have been collaborating with Dr. David Manlove and Dr. Péter Biró in relation to the NMSPD [National Matching Scheme for Paired Donation, now the UK Living Kidney Sharing Schemes]. Their matching algorithms have been used in order to construct optimal solutions to the datasets that we provide. Some of these datasets have encoded particularly challenging underlying problems, and the task of producing an optimal solution would have been highly complex without the assistance of these matching algorithms. We anticipate that this will be a growing issue as the number of people in the database increases over time.’

OUTCOMES

Transnational European collaboration (in which countries pool their datasets in order to obtain more transplants and better quality matches) will require the algorithms to be extended to ensure that they can cope with larger and more complex datasets

By optimising these PKEs and DPD chains, the algorithms have led to 752 actual transplants taking place between 2008 and 2017. Had NHSBT continued to use their pre-existing algorithm, which was only capable of identifying pairwise exchanges, it is estimated that 534 transplants would have gone ahead. Thus the 752 transplants that took place represents an increase of 218, or 41%, compared to the estimated number that would have occurred if the status quo techniques had continued to be used. According to NHSBT, each kidney transplant saves the NHS £240K over 10 years (based on a comparison with the cost of dialysis over that time period, and taking into account the cost of the operation itself ). This means that by enabling an increase of 218 new kidney transplants, the research has potentially saved the NHS around £52M over a 10-year period. In 2010, Rachel Johnson, the Head of Organ Donation and Transplantation Studies, NHS Blood and Transplant

FUTURE WORK

in order to obtain more transplants and better quality matches) will require the algorithms to be extended to ensure that they can cope with larger and more complex datasets. Work is ongoing to put in place the infrastructure to support these collaborations as part of the EU COST Action entitled ‘ENCKEP’ (European Network for Collaboration on Kidney Exchange Programmes), running from September 2016 to 2020, for which the author is Vice-Chair. The algorithmic objectives are also supported by the £800K EPSRC funded project entitled ‘IP-MATCH’ (Integer Programming for Large and Complex Matching Problems), joint with the University of Edinburgh, running from November 2017 to October 2020, for which Dr Manlove is the University of Glasgow Principal Investigator. David Manlove is a Senior Lecturer in Computing Science at the University of Glasgow. The work described above was an impact case study for the School of Computing Science’s REF 2014 submission. It has also been featured on the EPSRC website and as part of a BBC4 documentary on algorithms.

Transnational European collaboration (in which countries pool their datasets FOR FURTHER READING Biró, P., D.F. Manlove and R. Rizzi (2009). Maximum weight cycle packing in directed graphs, with application to kidney exchange programs. Discrete Mathematics, Algorithms and Applications 1: 499–517. Dickerson, J.P., D.F. Manlove, B. Plaut, T. Sandholm and J. Trimble (2016). Position-indexed formulations for kidney exchange. In Proceedings of EC 2016: the 17th ACM Conference on Economics and Computation, 25–42, ACM. Manlove, D.F. and G. O’Malley (2014). Paired and altruistic kidney donation in the UK: Algorithms and experimentation. ACM Journal of Experimental Algorithmics 19(2): article 2.6.

IMPACT | spring 2018

U N I V E R S I T I E S M A K I N G A N I M PAC T EACH YEAR STUDENTS on postgraduate programmes at several UK universities spend their last few months undertaking a project, often for an organisation. These projects can make a significant impact. This issue features reports of projects recently carried out at two of our universities: Warwick and Edinburgh. If you are interested in availing yourself of such an opportunity, please contact the Operational Research Society at email@theorsociety.com CHOICE MODELLING IN E-FULFILMENT (Yifan Cao, University of Warwick, MSc Business Analytics)

E-commerce is growing quickly in the UK; in fact, the UK has globally the third-largest B2C e-commerce turnover ($175bn in 2015). In particular, online grocery sales have shown double-digit annual growth for several years. The retailers are therefore facing an increasingly challenging e-fulfilment problem, especially because competitive pressure has forced almost all large UK grocery retailers to offer one-hour delivery time slots. Provision of such relatively narrow time windows comes with substantial fulfilment costs. Demand over a typical day is highly unbalanced so that fleet utilization likewise is unbalanced and therefore inefficient. To bring down delivery costs, retailers like Sainsbury’s are keen to look into customer choice behaviour of delivery time windows so as to gain insights on how are customers influenced by different delivery charges and slot availabilities. Yifan conducted an empirical study using online grocery sales data from Sainsbury’s to model customer choice

IMPACT © THE OR SOCIETY

behaviour in the context of delivery time slot selection. Specifically, she devised and estimated the parameters of a nested choice model that first considers the choice of the day of delivery, followed by choice of the time slot within a given day. The choice model has been validated using some holdout sample data and has been found to predict slot choice behaviour very well. The output of this work can be used in other management decision support tools to improve the efficiency of home delivery logistics. For example, the demand implications of offering, or of not offering, certain slots onto other slots can be predicted and used for planning purposes. Likewise, the choice model could be used to set delivery charges with the aim of smoothing demand over delivery time slots, which would lead to better fleet utilisation and ideally greener delivery operations. Joel Lindop, Online Forecasting & Optimisation Manager at Sainsbury’s,

London, stated that ‘Yifan’s project broke new ground for Sainsbury’s in providing a methodologically robust analysis of the interaction between the slots offered to customers and their likelihood of their placing orders. Her approach is being integrated within our in-house tools for the management of slot availability to optimise performance of the Sainsbury’s Online business. We thank Yifan for her valuable contribution.’ Yifan said that ‘taking the project for Sainsbury’s was a great challenge, yet the best decision I have made. The learning experience was very valuable and interesting with the help of my devoted and experienced supervisor.’ Her dissertation won the First Prize of the Consumer Data Research Centre (CDRC) Masters Research Dissertation Programme in 2017. Guy Lansley of the CDRC said ‘all the projects this year were really good so it was a particular achievement.’

L E V E R AG I N G P R E D I C T I V E A N A LY T I C S F O R C R I M E P R E V E N T I O N (Ritesh Kotak, University of Edinburgh, MBA)

If crime can be predicted using data, then an agency may be able to prevent it from occurring in the first place. The use of predictive analytics can be found in retail, finance and the health care sectors, but what about in policing for crime prevention? The objectives of Ritesh’s research was to answer three questions around the use of predictive analytics within policing: 1. What are the costs? 2. What are the benefits? 3. Do the costs outweigh the benefits? Overall, crime rates in Canada are decreasing, yet the data from a police agency in Canada showed that major violent crimes were in fact increasing. Concurrently, the clearance rate which illustrates the rate that crimes are solved is decreasing. This creates a concern of more violent crimes growing within the community which require a need to think outside the box on addressing these growing challenges. Leveraging the ongoing advancements in data analytics and machine-learning, a public safety organisation such as a police department may be able to effectively deploy resources to disrupt crime

patterns and thus prevent crime and future victimisation from occurring. The inspiration for this method comes from Mother Nature and how we predict natural disasters. Imagine visualising and modelling crime patterns to earthquakes and aftershocks. By allowing an incident to be classified as an aftershock linked to an earthquake, or in this case, a crime linked to a crime generator, the user can potentially forecast future incidents with greater accuracy. Using a similar concept, crime data was first analysed for repeats (same address) and near-repeats (same block). The surprising results showed that location had the greatest correlation with assault occurrences with just under two-thirds of all reported incidents. This was followed by both sexual assaults and break and enters at about a one-third of all incidents committed as either repeats or near repeats. Finally, the result for robberies was around 20% and no correlation was found for homicides. Given these finding, further research was conducted to calculate the additional costs incurred from these incidents from the perspectives of police agencies and the society. From these findings, which were

previously unknown, a financial analysis was conducted, where it was determined that a prediction accuracy rate of only 2–3% of the repeat or near repeat crimes would allow an agency to break-even in their investments. Police Chief Paul Martin from the Durham Regional Police Service (DRPS) says that, ‘Enhancing our ability to predict crimes by leveraging the next generation of predictive analytics shows promise in preventing repeat or near repeat crimes. Considering what we have known for years about the Problem Analysis Triangle (also known as the Crime Triangle) it should be no surprise that one of the three elements, being location, has a large influence on some crimes. Using the findings of this research, we will be further exploring predictive analytics within the DRPS and conducting a proof of concept in the near future.’ Predictive analytics within policing is in its infancy but already shows great promise. Early detection will lead to prevention. Predictive policing is a new opportunity for Canadian agencies to finally be more proactive and less reactive while working with all stakeholders to co-create public safety.

I M P A C T | s p r i n g 2018

21 21

MARTIN PITT AND KEN STEIN

THE NATIONAL HEALTH SERVICE is a complex and disparate organisation, and one in which resources are often stretched to the limit. Yet with drives for improved efficiency, NHS managers are often required to make vital operational decisions, which impact the way their

local service is delivered, and have direct consequences for patient care. An evidence-based approach is crucial to ensure small and large-scale changes have the intended outcomes, and a small team of operational researchers are leading this approach across the South West of the UK.

PENCHORD: HELPING THE NHS MAKE INFORMED DECISIONS USING O P E R AT I O N A L R E S E A R C H

Based at the Universities of Exeter and Plymouth, the Peninsula Collaboration for Health Operational Research and Development (PenCHORD) has, over the last eight years, been working with the NHS to make simulation and modelling techniques a key part of the decision-making process.

PenCHORD has been working with the NHS to make simulation and modelling techniques a key part of the decision-making process

A DIVERSE TEAM

PenCHORD is part of the Collaboration for Leadership in Applied Health Research and Care South West Peninsula (PenCLAHRC), a National Institute for Health Research initiative launched in 2008. With a mandate to put research findings into practice, PenCLAHRC enlisted the help of Professor Martin Pitt, who had previously established the UK Network for Modelling and Simulation in Healthcare (MASHnet) in 2005. Martin was keen to build on the motivation of MASHnet, which explored and discussed how to improve the use of operational research tools to enhance healthcare. The PenCHORD research initiative was established in 2008 to provide a platform to enact and implement many of the aspirations articulated through MASHnet. PenCHORD now brings together researchers from a range of diverse backgrounds spanning not just operational research but computer science, psychology and pharmacology, and specialists in optimisation, simulation, behavioural modelling and health economics.

SPREADING THE WORD

In early efforts to reach out to healthcare managers, the team recognised that operational research had an identity problem. They found that most people were unfamiliar with the discipline, and that it was perceived as being esoteric and technical. They set about addressing these misgivings through a series of free-of-charge showcasing events, workshops and one-day training sessions. These courses are designed for those wishing to learn how to build simple but effective predictive models for their own organisations. In addition, regular seminar events are run for all those working in and around the NHS in the region to hear more about how operational research can help improve services. These sessions provided insights into the real-world application of modelling and simulation, and encouraged attendees to begin formulating their own questions. Eight years on, the team has worked with all of the trusts across the South West and is helping to change the way decisions are made.

IMPROVING STROKE TREATMENT

Time is one of the most important factors when treating stroke patients and, in some cases, administering clot-busting medication can limit brain damage and reduce the likelihood of disability. Yet working out whether a person with stroke is suitable to receive this treatment can be complicated and time-consuming. To find ways to speed up this process, PenCHORD partnered with the Royal Devon & Exeter (RD&E) Hospital and the South Western Ambulance Service to create a detailed working model of the path that patients

take when they have a stroke, replicating the journey from onset of the condition right through to the point of treatment. They included the roles of paramedics, nurses, emergency department clinicians, the acute stroke team and the radiology department, and then used this model to test how changes in the system might affect the delivery of care. With several new initiatives established, the team implemented their findings at the RD&E Trust. The results of the project were remarkable. Within 18 months of implementation the number of patients with acute ischemic stroke receiving medication more than doubled, increasing from 4.7% to 11.5%. The average time patients had to wait for this treatment also dropped substantially, coming down from 58 to 33min.

Within 18 months of implementation the number of patients with acute ischemic stroke receiving medication more than doubled

And the benefits didn’t stop there. Now, four years after the new processes were put in place, treatment rates have increased threefold from their original levels, up to 14%. For over 600 people who are admitted with an acute stroke every year, the RD&E is now matching or beating the performance of the big urban stroke centres elsewhere in the UK and the world. ‘We had very little understanding of computer modelling but it’s been absolutely amazing to be part of the process, changing practice in such a

IMPACT | spring 2018

short space of time and making a tremendous difference to patient care’, says Nicki Chivers, Senior Stroke Nurse at Royal Devon and Exeter Hospital.

BEATING BLADDER CANCER

An analysis of data six months after the changes revealed significant reductions in waiting times, speeding the waiting time for treatment up by 35 days in some cases

IMPACT | spring 2018

An analysis of data six months after the changes revealed significant reductions in waiting times, speeding the waiting time for treatment up by 35 days in some cases. The simulation and analysis were led by Dr Daniel Chalk: ‘This was a fascinating project for us and one which had translated beautifully from computer simulation to proven improvements in care for patients. By working closely with clinicians “on the ground” we have come up with solutions that work for both them and their patients.’ With 10,000 or so people diagnosed each year with bladder cancer in the UK, the team are hoping their success will be replicated in other Trusts across the country.

CREATING CAPACITY

PenCHORD’s collaborations with local healthcare Trusts have helped to highlight the powerful role operational research can have in the NHS, but the team’s long-term goal has always been to build self-sustaining modelling capacity within the organisations themselves. To that end, in 2016 the team launched a 12-month secondment

programme designed to train NHS staff in the use of simulation and modelling techniques. The Health Service Modelling Associates (HSMA) programme released participants for one day a week to work with the PenCHORD team, and the first wave of the programme saw staff from six NHS organisations across Devon and Cornwall sign up to take part. Mentoring and support was delivered via a two-day residential training course, monthly learning-set meetings, weekly support, and a final showcase event. Each participant tackled an important problem facing their Trust and one year on, the individuals and institutions involved are already seeing significant benefits. In one example, Northern Devon Healthcare NHS Trust improved both wait times and admission rates in Accident and Emergency. Nic Harrison, Senior Analyst and HSMA trainee, built a detailed model of the way patients flow through their emergency department so he could experiment with various configurations. The final optimised solution allowed the Trust to meet crucial targets and receive a financial reward that could then be reinvested in improving patient care.

It can often be hard to spot the elements that make a system inefficient, and healthcare pathways that have grown responsively can be difficult to unravel. In 2016, the Royal Cornwall Hospitals Trust identified possible improvements to their bladder cancer pathway, but lacked the skills to explore these changes without a trial-and-error approach. So they teamed up with PenCHORD to try and bring the average time patients wait to be referred for definitive treatment down from its 90 day average. Using two years’ worth of data, the group created a detail process map and simulated the flow of patients through the system. They worked with consultants and urologists to identify practical changes that might reduce delays, and tested their impacts using the model. With two key reconfigurations made at no additional cost, the simulation predicted that the average wait time for treatment could be reduced by several weeks. This evidence was so compelling that the department swiftly re-wrote their bladder cancer patient protocol, implementing it just 24 hours later.

Northern Devon Healthcare NHS Trust improved both wait times and admission rates in Accident and Emergency

‘We’ve been able to reap the benefits of this way of working in areas outside of the emergency department, extending our work to impact decisions on ward numbers, cardiac treatment and eye care’, says Nic Harrison, Principal Analyst, Northern Devon Healthcare NHS Trust. In addition to the training in operational research techniques, the Associates have developed skills around working with stakeholders, structuring problems, and communicating project findings. They are now acting as ambassadors for the use of modelling in their organisations.

As the NHS experiences growing pressure to deliver services in the face of increasingly limited resources, the role that operational research can play in solving the day-today problems becomes ever more important

ongoing success in increasing engagement with healthcare professionals.

CONCLUSION

As the NHS experiences growing pressure to deliver services in the face of increasingly limited resources, the role that operational research can play in solving the day-to-day problems becomes ever more important. We want to ensure that our research is making a real difference to NHS services and ultimately patient care. Modelling techniques can be an immensely powerful tool in the decision-making process and we’re always eager to hear from healthcare workers who might have a problem that we can help with, regardless of its size or scope. Martin Pitt is the Director of PenCHORD, whose main interest is the effective implementation of modelling, simulation and analytic solutions to inform health policy and the evaluation of its impact. A particular focus is the communication of information and visualisation techniques to facilitate understanding across organisational and cultural boundaries. He completed his doctorate at the Medical Informatics Group at Manchester University

developing an interactive multimedia system for use by healthcare consultants. In 1998 he joined the University of Exeter as a lecturer in media computing before joining the Peninsula Technology Advisory Group (PenTAG) in 2003 where he helped develop a range of economic models in the context of health technology assessment. Ken Stein is the Deputy Director of the NIHR CLAHRC South West Peninsula (PenCLAHRC), where he has Executive responsibility for the PenCLAHRC Evidence Synthesis Team and PenCHORD. He was a member and Vice Chair of a NICE Technology Appraisals Committee for 15 years. He graduated in Medicine from University of Bristol in 1987, then trained and worked as a GP in Australia and Hampshire before specialising in public health medicine in Southampton, where he became the Deputy Director of the NIHR Evaluation, Trials and Studies Coordinating Centre (NETSCC), Health Technology Assessment (HTA) programme. In 1999 he began work as a Consultant in Public Health Medicine at North and East Devon Health Authority, then Director of Public Health for Mid Devon Primary Care Trust. He was a founding Director of the Peninsula Technology Assessment Group (PenTAG), and appointed to a Chair in Public Health in 2007.

FIND OUT MORE

Building on the success of the 2016 round and the strong national interest in the initiative, PenCHORD have recently launched the 2018 iteration of the HSMA programme, in partnership with South West Academic Health Science Network and their Regional Intelligence Analysts Network (RIAN). The 2018 programme attracted almost 50 applications, seven times as many as in 2016. From this 26 Associates have been recruited, demonstrating PenCHORD’s

PenCHORD: http://clahrc-peninsula.nihr.ac.uk/penchord Health Service Modelling Associates (HSMA) programme: http://clahrc-peninsula.nihr.ac.uk/ health-service-modelling-associates-programme 2016 HSMA programme website: https://health-modelling.org/ NIHR CLAHRC South West Peninsula (PenCLAHRC): http://clahrc-peninsula.nihr.ac.uk MASHnet: http://mashnet.info/ South West Academic Health Science Network: https://www.swahsn.com/

IMPACT | spring 2018

HOWARD TURNER, ANDREW REEVES AND TARA ROWE

CHARITYWORKS is a paid 12-month graduate scheme for the UK social sector. It aims to attract, grow and retain future managers and leaders, to increase the reach, quality and impact of its services and improve the lives of the people they serve. It collaborates with 75 charitable organisations. The trainees undertake full-time jobs in a partner charity or social housing association. It has an acclaimed

leadership development programme, and in seven years it has grown from 3 trainees to 120. Charityworks turned to the OR Society’s Pro Bono Scheme, and Andrew Reeves and Howard Turner from the Government Operational Research Service helped them streamline their recruitment process and develop an automated system for matching successful candidates to posts.

Photo by Elliott Shephard © Charityworks

O P E R AT I O N A L R E S E A R C H IN SUPPORT OF C H A R I T YWO R K S

FIGURE 1 STAGES OF RECRUITMENT PROCESS

CANDIDATE SELECTION

The first question to tackle was to do with the selection process. This had grown to be quite complex, as shown in Figure 1. This shows a process where candidates start by registering and submitting an application form. Then they undergo numerical and verbal reasoning tests. If their scores on these tests are satisfactory, their application forms are marked against a number of criteria. On the basis of this information, some are invited to assessment centres, where they have an interview and take part in group and written activities. For the candidates successful at this stage, there is a final selection day with an interview and presentation. After that there follows the process of matching successful candidates to posts, which we tackled in the second part of the study described below. Charityworks wanted to know whether it would be possible to simplify the system, in particular: What would the effect be of replacing/removing certain tests? Which tests indicate a good trainee? How discriminatory are the psychometric tests?

The first of these questions was at first sight quite hard to tackle. We only had a small amount of information on the performance of successful candidates (and none at all on the unsuccessful ones). We also did not know what weightings had been used for the different tests in past years. For instance, some tests were marked on a scale of 1–10 and some on a scale of 1–25. If one just added the scores together the latter would receive more weight than the former. Alternatively, one could use the percentage of the possible maximum mark for each test, in which case they would all receive equal weight. We agreed to use a conventional weighting based on what seemed sensible as the baseline. We modelled the previous year’s candidate scores using that weighting and an alternative with one or more elements removed. Here, elements can cover either individual tests or combinations of tests or such things as group activities. Then we compared the results. If the set of successful candidates was similar with and without a particular element, then it could possibly be

DATA ISSUE

dispensed with. We found that removing the Selection Day, which comprised the three elements of Placement, Interview (Selection Day) and Presentation, meant that 13 different candidates would be promoted into the passing group of 120 successful candidates. Our recommendation was: On this analysis, dispensing with the Selection Day would lead to a change of 13/120 or 11% in the successful candidates. On that basis, it may be worth considering whether its retention is worthwhile. We also carried out some analysis of the effectiveness of the different test elements in predicting candidate performance. This was based on a small number (19 in all) of successful candidates adjudged to be especially high or low-performing in their posts. It showed for instance that for a test identified as Question 1 the high-performing trainees had a higher mean score than the low-performing ones. The difference was highly significant (t < 0.01), and so this test was taken as a good predictor. We did come across issues with data quality that required some work to resolve, as shown in Table 1.

SOLUTION

Undocumented (it was hard to tell

Ask Tara about it, or deduce from relation to known data items

what it was) Missing

It was sometimes possible to reconstruct missing data from what was presented; otherwise overall averages could be used

Had original data mixed with derived

Put original data and derived quantities in separate spreadsheets

quantities Had derived quantities calculated

Recalculate

incorrectly Corrupt

Restore by hand

TABLE 1 ISSUES WITH DATA QUALITY AND SOLUTIONS APPLIED

IMPACT | spring 2018

FIGURE 2 MANUAL PROCESS OF MATCHING SUCCESSFUL CANDIDATES TO POSTS

Data on candidate scores posed particular problems where some quotients involving recurring decimals had been converted into millions, but not consistently. These millions would then sum to an apparent answer of 300 or so. How 80/3 apparently became 2,666,666,667 in some places and 26,7 in others was a special mystery. There seemed to be no easy way of dealing with these millions so this had to be done by hand. As for outcomes, since Numerical Reasoning had a perverse sign (the low-performing trainees scored better), it was replaced by a bespoke Situational Judgment Test. Verbal Reasoning was retained as a good predictor, but moved to a new test supplier. This was because the original test was unfavourable for candidate diversity. The Selection Day was abolished, since it took up resource and was not that informative.

This gave Charityworks more certainty about decisions they had been grappling with. The recommendations promised significant savings in volunteer time

This part of the project gave Charityworks more certainty about decisions they had been grappling with. The recommendations promised significant savings in volunteer time.

CANDIDATE PLACEMENT

The other arm of the study addressed how candidates who had passed the selection process could best be fitted to available posts.

the study addressed how candidates who had passed the selection process could best be fitted to available posts

As the number of successful candidates and posts had increased year-on-year, Charityworks had found it increasingly difficult to match people to posts by hand. In 2015, the process of matching 120 candidates to posts by hand took 3/4 weeks of FTE effort. They were looking for an automatic process that would do this for them. This was an especially complex problem for them. There was a huge diversity of roles and causes. Candidates and host organisations had different priorities and preferences. The host organisations would also drop out or change their requirements throughout the process. Figure 2 shows the various stages of the process, as it was implemented by hand. This is a classical O.R. problem known as the Assignment Problem. The general question is of assigning a number of people to a number of

jobs so as to maximise overall satisfaction. One approach to this is that all of the candidates express a preference (1st down to 120th in this case) for each of the jobs and the employers do the same for the candidates. It is possible to derive an optimal matching this way, but here it would be impractical to ask each candidate to rank 120 posts; and similarly for the employers. The alternative is to consider the requirements expressed by the candidates and employers as a number of constraints. Then each matching is assigned a cost depending on how many constraints are violated and how badly. The art is to find the assignment (set of matchings of all candidates to all posts) which has the minimum overall cost. It was not trivial to define the constraints regarding which candidates could occupy what posts, or to produce an assignment on that basis. There was a mixture of ‘hard’ and ‘soft’ constraints. A hard constraint might be that a candidate could only take a post in a particular geographical area. A soft constraint might be that s/he would rather work on particular issues.

the number of such sets of matchings is vastly greater than the number of elementary particles in the observable universe

IMPACT | spring 2018

FIGURE 3 THE PROTOTYPE MODEL

With 120 candidates and posts, the number of such sets of matchings is vastly greater than the number of elementary particles in the observable universe, a result of “combinatorial explosion. (In mathematical terms, the number of possibilities is 120! or 120 factorial, which would be about 7 followed by 198 zeroes if written out.) It did, however, prove possible to use a reasonably simple approach, in the form of the so-called Hungarian algorithm, to tackle this problem. The basis is that one starts with the ‘cost matrix’ (set of costs) for a trial assignment and then arrives at the optimal assignment by a relatively small number of elementary operations. By employing an effective method of constant improvement, the algorithm only needs to consider a very small proportion of the possible assignments to reach the best solution. This was a practical solution that could find the optimum assignment in a reasonable time. It could be

IMPACT | spring 2018

implemented in Excel/VBA, which was also consistent with existing Charityworks IT practice. Tara felt it was something that Charityworks could, if necessary, adapt themselves in the future. So we began the process of interactive model development. It proved to be quite straightforward to produce a prototype which would match people to posts under a modest number of hard constraints. Producing an assignment typically took less than a minute. Figure 3 illustrates the output of the prototype model, using some fictitious charity and candidate names. We then continued to try to produce a model for routine use. In retrospect, it is not clear that this was a decision that was taken explicitly, but it certainly happened. While production of a prototype had gone smoothly, adding in different tiers of (especially soft) constraints seemed to give rise to new difficulties at every stage. Perhaps this

was yet another example of a ‘combinatorial explosion’. We had a number of meetings trying to bring this model to a conclusion. In the end we decided to produce a stable version of the model with those constraints we had been able to implement successfully. After completing of work on this second part, we received some further input from leading figures in the UK O.R. Community. This fed into further developments. First of all, an MSc student project took the model further and then Charityworks produced a brief for a professional software developer.

CONCLUSIONS

The first part of the project (on analysing the performance of different components of the selection process) went much as planned. It produced results that at the very least gave Charityworks more confidence in what they were planning to do.

The second part of the work provided a solution at a working prototype level and showed how an operational solution could be devised. It also took up twice as much time as had originally been planned! In retrospect, it would have been useful to make absolutely sure we had exhausted all the avenues for finding an off-theshelf solution. The demarcation between problem-solving and providing an operational software product could also have been made clearer. Reflecting on the experience, it was clear that the project had highlighted the need for Charityworks to improve their data systems. It is an important point that where data is not used for actual decisionmaking it will often be of poor quality. Similarly, where many different people work on a particular task it is especially important to have effective ways of storing, checking and cataloguing data.

where data is not used for actual decision-making it will often be of poor quality

Andrew Reeves joined GORS in 2014 after completing a PhD in Mathematics. During his time with GORS, he has worked on policy analysis and in support of HR.

With regard to candidate matching, everyone felt that it would have been good to be clearer about the scope of the work–and indeed to be clearer on whose task it was to define this scope. The particular value of O.R. was in structuring and solving problems. Software implementation would be at a prototype level. A software developer could then implement a routine solution if necessary.

Tara Rowe joined Charityworks after completing the graduate programme in 2012. During this period of this project she was Programme Manager, managing candidate selection process, quality and impact, as well as another charitable programme for disabled students. She now works as an analyst with Social Finance, another not for profit organisation.

Howard Turner joined the Government Operational Research Service (GORS) in 1990. His time has included two periods with overall responsibility for GORS recruitment. At present he leads a number of projects in support of Human Resources.

For more information about Pro Bono O.R. please contact project manager Hope Meadows at hope.meadows@ theorsociety.com. Alternatively, please visit http://www.theorsociety.com/ Probono.

IMPACT | spring 2018

LUCA GRIECO AND MARTIN UTLEY

POWERED RESPIRATOR PROTECTIVE SUITS (PRPS) are special suits worn by trained operators when carrying out decontamination of people exposed to chemical, biological, radiological, or nuclear (CBRN) materials. In 2014, the National Health

Service (NHS) England faced the problem of establishing how many of these suits each ambulance service and each hospital with an emergency department in England should have. The UK Department of Health (DH) commissioned the Clinical Operational

I M P R OV I N G E N G L A N D ’ S R E S I L I E N C E TO I N C I D E N T S I N VO LV I N G RELEASE OF HAZARDOUS M AT E R I A L S

Government is responsible for making recommendations to service providers who are then responsible for decisions about buying and managing resources. In this context, NHS England were to give guidance to ambulance services and hospitals with an emergency department in England about the number of Powered Respirator Protective Suits (PRPS) to have in stock. This had to be done prior to current stock going beyond their expiry date.

AN ANALYTICAL FRAMEWORK TAILORED TO THE NATIONAL CONTEXT

Research Unit (CORU) at University College London (UCL) to address this question, with NHS England as the client for the work. CORU developed an analytical framework to give NHS England a means of ensuring a given degree of resilience to a set of potential CBRN accidents. Dr. Peter Grove, Senior Principal Public Health Analyst at the UK Department of Health said ‘CORU provided DH with a responsive Operational Research facility to support health protection policy for 10 years. The piece of work on powered respirator protective suits was particularly valuable and, to me, is a prime example of academic Operational Research supporting decisions of national importance.’

DEALING WITH HAZARDOUS MATERIAL RELEASES

Incidents involving release of CBRN materials can have a significant social and health impact. When caused by human error, technological failure or, for example, extreme weather events, these are commonly referred to as ‘HazMat events’. Such accidents, as well as malicious incidents (criminal or terrorist

acts), have the potential for significant human losses and environmental damage. NATO’s guidelines for first response give top priority to minimising the number of human deaths. Healthcare workers are required to establish decontamination and triage areas and to carry out decontamination procedures in order to end casualties’ exposure to the hazardous substance as soon as possible, and prior to further clinical treatment. Decontamination is also fundamental to prevent the spread of toxic substances to other people/areas.

The analytical framework CORU developed was influenced not just by the intrinsic characteristics of the problem at hand but also by the client and project sponsor perspectives on the information available about potential HazMat events and by the nature of the decisions to be made. Moreover, lack of detailed information about event locations and precise estimates of event likelihood led CORU to follow a precautionary approach for determining the allocation of protective suits. The methodology adopted consisted of four steps as shown in Figure 1.

The piece of work on powered respirator protective suits was particularly valuable and, to me, is a prime example of academic Operational Research supporting decisions of national importance

Step 1: selection of HazMat events

In September 2013, the UK Government reviewed the requirements for UK responders (fire brigades, police, healthcare services, military forces) in respect of protective equipment for use in CBRN and other contaminated environments. In the UK, the

Descriptions of possible HazMat events and their impact (health, social, economic and environmental effects) are determined centrally in the UK National Risk Register. In each of the 39 police areas in England, a Local Resilience Forum (LRF), formed by key emergency responders, is required to select for a Community Risk Register (CRR) those events they consider relevant to them locally and attribute to each one of five likelihood levels of the event happening in the area within the next five years. Following advice by project sponsor and client, CORU took the set of CRRs for

IMPACT | SPRING 2018

simple mathematical function was used to model this feature, with the relevant parameters (length of time window and peak time) agreed for each type of event with experts from the UK Department of Health and NHS England. Step 3: single-event demand estimation

FIGURE 1 OVERVIEW OF THE METHODOLOGY ADOPTED BY CORU TO INFORM THE STOCK LEVELS OF PROTECTIVE SUITS TO BE HELD BY EACH AMBULANCE SERVICE AND HOSPITAL WITH AN EMERGENCY DEPARTMENT IN ENGLAND

England as the starting point for the analysis in order to align the work with the existing relevant decision processes. From the CRRs a list of HazMat events potentially requiring decontamination was extracted. In the absence of more detailed information, the ‘number of casualties’ or ‘hospital admissions’ reported in the event descriptions were used as estimates of the number of people requiring decontamination. Local assessments of event likelihoods were also retrieved from each CRR. Step 2: mathematical characterisation of HazMat events

A number of factors can determine the proportion of contaminated patients that remain at the site of a particular HazMat event. For instance, an immediately detectable event in a place where safety procedures are well established (e.g. an accident at a chemical plant) would most likely involve only ambulance services, as all the contaminated people would be isolated and waiting for the decontamination units to be set up. On the contrary, a silent release of toxic substance in a public place would

IMPACT | SPRING 2018

probably be detectable only after many contaminated people have already left the scene, and consequently a proportion of casualties would need to be treated at emergency departments. Experts from the National Ambulance Resilience Unit (NARU) advised on the proportions of ‘on-site’ versus ‘self-presenting’ casualties associated with each of the HazMat events identified from the CRRs. While all ambulance casualties are already on site when decontamination procedures start (therefore the number of people to treat is already known), self-presenting casualties attend hospitals individually with an arrival rate that will change over time from the event. The rate at which patients arrive at a hospital following an event strongly depends on when they start to recognise the symptoms, which may vary based on the specific contaminant, or when they become aware of the incident (e.g. through mass media). It was assumed that, following a HazMat event of given characteristics, patients would arrive at the ED for decontamination during a time window with some peak time at which the arrival rate reaches its maximum. A

The core of the analytical framework consisted of estimating the demand for protective suits for each healthcare service in response to single events occurring in a given region, based on the expected number of casualties and on the characteristics of the decontamination procedures carried out. Decontamination of casualties exposed to hazardous materials is carried out in decontamination units (usually tents) that are assembled as required. Official guidelines regarding the deployment of decontamination units in the UK provided CORU with relevant information to build mathematical models capturing salient aspects of the decontamination process.

The core of the analytical framework consisted of estimating the demand for protective suits for each healthcare service in response to single events occurring in a given region

In order to determine single-event demand estimates, CORU modelled the decontamination process as a queueing system. Queueing theory models are very useful for studying the dynamics of processes characterised by a service to be delivered to customers. Different assumptions can be made around demand features (e.g. arrival pattern of contaminated casualties at an emergency department) and service features (e.g. time needed to decontaminate a

casualty, number of workers carrying out decontamination procedures). A different mathematical queueing model was built for ambulance services than for emergency departments, given the different ways patients present to them for decontamination. Decontamination by ambulance services is characterised by the fact that all contaminated people are at the site of the HazMat event. This process can be represented using a clearing model, that is, a queueing system with the initial queue sized as the number of on-site casualties associated with the given HazMat event. A simple algorithm enabled estimation of the number of decontamination sessions needed, and thus the requirement for protective suits. A more complex model was developed for emergency departments. In this case, staff members may not be aware of the type and scale of an HazMat event when contaminated people start self-presenting. A plausible assumption was made that the capacity of the system would be modified during the incident depending on the current number of patients in the queue. An existing method approximating the dynamics of the system was embedded in an algorithm developed by CORU enabling estimation of the number of decontamination sessions needed in response to a given event. Step 4: informing national guidelines

The last step of the work consisted of translating the numbers obtained from the single-event analysis into stock levels of protective suits to be recommended by NHS England for each ambulance service and each hospital with an emergency department. A ‘plausible worst-case scenario’ approach was agreed with the client. With such an allocation strategy, the required stock of protective suits to be held by a healthcare service responsible

in a given region corresponds to the maximum single-event estimate obtained for that healthcare service across all events with likelihood level equal to or above a given threshold in that region. Note that ‘worst’ here does not necessarily correspond to ‘largest number of exposed people’, but rather to ‘largest number of suits needed’ to decontaminate the exposed people.

This project provided us with an evidence base, incorporating local risk assessments, that we could use in guidance to ambulance Trusts and hospitals about the purchase of this important equipment, directly informing decisions valued at millions of GBP

However, this does not account for the possibility of a second event occurring in the region before the appropriate stock of protective suits is replenished. Healthcare services might not be able to respond to that second event, and to a third, and so on. To deal with this, CORU implemented the whole analytical framework for different threshold levels of likelihood and shared the corresponding ‘worst’ single-event stock estimates with NHS England. Guidelines were drawn using these results as ‘building blocks’, that is, stocks of protective suits were determined by summing up demand estimates obtained for combinations of worst-case scenarios of interest. Stephen Groves OBE, National Head of Emergency Preparedness, Resilience at NHS England said ‘This project provided us with an evidence base, incorporating local risk assessments, that we could use in guidance to ambulance

Trusts and hospitals about the purchase of this important equipment, directly informing decisions valued at millions of GBP. Luca and Martin worked in close collaboration with us to make sure they understood our needs and that we understood the analytical approach and how to use the results.’ Luca Grieco (l.grieco@ucl.ac.uk) is a Research Associate at CORU where he applies quantitative methods to support health protection policy and to improve performance of healthcare systems. Before joining CORU, Luca graduated in Industrial Engineering at Sapienza University of Rome and then obtained his PhD in Genomics and Bioinformatics at Aix-Marseille University dealing with machine learning and dynamical modelling approaches to identify targets for cancer therapy. Martin Utley (m.utley@ucl.ac.uk) joined CORU in 1996 having gained a PhD in high-energy physics. From 2007 to 2017 he led the responsive Operational Research facility that CORU provided to the Department of Health to support Health Protection Policy. His other academic and consultancy work focusses on helping clinicians, service managers and analysts in their work to improve services. This article is based on independent research commissioned and funded by the NIHR Policy Research Programme (project reference 027/0085). The views expressed in the publication are those of the author(s) and not necessarily those of the NHS, the NIHR, the Department of Health, arms length bodies or other government departments. The full account of this work is due to appear in the International Journal of Disaster Risk Reduction - see https://doi. org/10.1016/j.ijdrr.2018.02.036

IMPACT | SPRING 2018

MEASURING EFFICIENCY T H R O U G H DATA E N V E LO P M E N T A N A LYS I S EMMANUEL THANASSOULIS AND MARIA CONCEIÇÃO A. SILVA

KEY PERFORMANCE INDICATORS (KPIs) are typically deployed in the ever present imperative to improve productivity. They are intended to enable comparisons on performance and target setting. However, KPIs are ratios of one deliverable (e.g. patients treated in a hospital) per unit of resource (e.g. consultants), so that a single KPI rarely captures the full complexity of a production system. Data Envelopment Analysis (DEA) is a method which enables comparisons where units use multiple incommensurate resources (‘inputs’) to deliver multiple incommensurate outcomes (‘outputs’), to yield a single measure of overall performance. Along with the measure, DEA also yields targets for performance, any gains realizable through changes in scale size and/or mix of resources used, identification of best practice and benchmark units. Here we outline how DEA works and indicate some key areas of its use. The origins of DEA can be traced to the seminal paper by Charnes and his co-authors in 1978. They operationalised, through Linear Programming (LP), the notion of using empirical data from operating units to measure their comparative performance.

A convenient point of departure for understanding how DEA works is to see it as an extension of KPIs

HOW DOES DEA WORK?

A convenient point of departure for understanding how DEA works is to see it as an extension of KPIs. For example, if we want to compare hospitals on their perinatal care the outputs capturing activity levels may be number of normal baby deliveries, number of babies delivered to intensive care, number of pregnancies cared for, etc. The inputs might be number of consultants, paediatricians, midwives, nurses etc.

DEA constructs a composite KPI as the ratio of the weighted sum of outputs to a weighted sum of inputs

The idea of the outputs is to capture as comprehensively as possible the deliverables from the units and the idea of the inputs is to capture as comprehensively as possible all resources used and any contextual factors that affect the outcomes being delivered by the units being compared. So essentially, DEA constructs a composite KPI as the ratio of the weighted sum of outputs to a weighted sum of inputs. You might ask what is so special about this and how does all the other information such as performance targets, economies of scale etc. noted above come in? The answer lies in how the weights assigned to inputs and outputs are arrived at.

The weights are arrived at by means of a Linear Programming (LP) model. The model identifies weights which would be best for each unit being assessed in the sense of maximising its efficiency rating. The only restriction is that those same weights when applied to the input and output levels of the rest of the operating units should not permit any one of them to attain an efficiency rating greater than 100%. Through determining an optimal set of weights for each unit in turn we respect its operational autonomy. The unit is free to choose the weights that best favour its input-output levels. If, in this scenario, any units other than the one being evaluated attain a better ratio of sum of weighted outputs to inputs then the unit being evaluated does indeed

IMPACT | SPRING 2018

have room for improvement, relative to ‘benchmark’ units attaining top efficiency with the weights favouring the unit being assessed. Notice also that the measure of efficiency we derive is a summary measure capturing all inputs and all outputs and without the need for the evaluator to impose subjective weights on individual inputs and outputs.

THE CONCEPT OF TECHNICAL EFFICIENCY

The concept of DEA efficiency can also be approached through a ‘production space’ route which has equivalence to the notion of efficiency as a ratio of sum of weighted outputs to inputs. This ‘production space’

conceptualisation is illustrated in Figure 1 which depicts 7 power generating plants (A, B … H), in terms of the amount of each input (Labour hours, Installed capacity) they use per 1000 MWh of power. (Labour, capacity and costs on the graph are for illustrative purposes only.) The DEA efficiency of a power plant such as B is the fraction to which its labour and capacity per 1000MWh can be reduced. To determine this fraction, we assume that observed plants can in principle be combined in whole or in part to create benchmark plants. Then we can compare B with a suitable benchmark that can deliver the 1000MWh of B but using as low a fraction of B’s labour and capacity as possible. We need to determine

multipliers, one for each observed power generating plant, to combine them so that a suitable benchmark for plant B can be identified. Such multipliers are identified using LP. For example, solving the LP model in respect of plant B in Figure 1 we identify multiplier 0.45 for E and 0.65 for F, which yield a benchmark at B′. That is using 65% of the power, labour and capacity of plant F and 45% of those of plant E the resulting aggregate power, labour and capacity levels lead to a virtual plant at B′. The levels of capacity and labour at B′ are by design to the same ratio as those of plant B but they are lower in absolute terms. They reflect the proportion therefore to which its labour and capacity per 1000MWh can be reduced. Extending this notion across all observed plants leads to benchmarks which would lie on the solid line AFED. This is referred to as the ‘efficient’ part of the frontier. Real and virtual plants on the efficient frontier cannot decrease one of labour or capacity per 1000MWh without increasing the other. The observed plants A, F, E and D which map out the efficient part of the frontier are referred to as ‘efficient peers’. For example plants E and F are efficient peers to plant B because they give rise to the virtual plant at B′ which is the benchmark for the efficiency rating of the plant at B. E and F are chosen by the model as the closest efficient plants to B in terms of ratio of capacity to labour. This in turn makes plants E and F probably the best efficient peers from which B can draw advice to improve its performance. Plant C, though it lies on the frontier, is not efficient. It has potential to reduce labour without raising capacity per 1000MWh. Thus, not all frontier plants are fully efficient.

Plants in the interior of the production space are not efficient since both labour and capacity per 1000MWh can be reduced. Their distance from the frontier (as that between B and B′) conveys a measure of how much their labour and capacity per 1000MWh can reduce. This measure does not necessarily reflect the full scope for reducing the aggregate cost of labour and capacity per 1000MWh. For this reason, it is known as ‘technical efficiency’.

THE CONCEPT OF COST EFFICIENCY

Where we have input prices (or output prices for that matter) we can assess not only technical but also ‘cost efficiency’. In such a case we would measure the distance of each unit from a minimum cost frontier. For example, if with reference to Figure 1, reproduced as Figure 2, the unit price of labour is £20,000 and that of capacity £50,000, then the ‘cost frontier’ would be the ‘isocost’ line corresponding to aggregate

FIGURE 1 PRODUCTION SPACE ILLUSTRATION OF DEA

FIGURE 2 GRAPHICAL REPRESENTATION OF ISOCOST

IMPACT | SPRING 2018

FIGURE 3 CONTRASTING CONSTANT AND VARIABLE RETURNS TO SCALE TECHNOLOGIES

cost of 8.5 (£85,000) for the labour and capacity needed to deliver 1000MWh as depicted in Figure 2. Only plant D of all potential plants has labour and capacity levels meeting the aggregate cost of £85,000. The location of the cost frontier depends on the shape of the technical frontier and the slope of the isocost lines, which in turn depend on the ratio 2:5 of the unit prices of labour and installed capacity. So we now have two measures of the efficiency of Plant B. In technical terms one is OB′/OB, which takes unit B to location B′. However, in cost terms the unit would still have potential to save. That is reflected in the distance B″B′ and captured in the ratio OB″/ OB′. OB″/OB′ is known as the ‘allocative efficiency’ of unit B. This saving can only be achieved if B replicated the labour, capacity and power of plant D. It is worth noting that the cost efficiency of B can be seen as a product of technical and allocative efficiency,

OB�� ∕OB = OB�� ∕OB� × OB� ∕OB

IMPACT | SPRING 2018

ESTIMATING POTENTIAL GAINS THROUGH SCALE SIZE CHANGES

We can understand scale size effects on productivity with the aid of Figure 3, where the observed units are depicted by . The efficient frontier under constant returns to scale (CRS) is simply the line passing through C. That is, under constant returns to scale one assumes that the maximum productivity (i.e. ratio of output to input) found at C is replicable whatever the level of inputs and outputs. If we cannot maintain this assumption of proportionality between inputs and outputs, then maximum productivity will depend on the scale size and we say that variable returns to scale (VRS) apply. Under VRS, the line through C in Figure 3 is not the proper frontier because it implies constant proportionality between input and output levels irrespective of scale size. Instead, under VRS the frontier is ECFD. This reflects virtual units (benchmarks) created as unscaled averages of observed units; that is, using multipliers for observed units

that add up to 1 and not scaling up or down observed units. Each production unit can now be assessed under an assumption of CRS and VRS. For example, if we measure efficiency in the input orientation so that output is held constant, then unit A’s efficiency under CRS is OA*CRS/OA while under VRS its efficiency is OA*VRS/OA. The distance between the CRS and VRS frontiers at the output level of unit A is (OA*CRS/OA*VRS). It is known as ‘scale efficiency’. It captures the amount by which unit A can reduce input even after it reaches point A*VRS at which it is fully efficient under VRS. This potential gain in input reduction will only be possible if unit A attains the same level of productivity (output to input ratio) as the most productive unit observed, which is unit C. This would mean unit A has to reduce its scale size to that of unit C. Unit A is therefore said to operate under decreasing returns to scale. DEA software (e.g. www. deasoftware.co.uk) automatically produce the scale efficiency of each unit, identifying whether it is reflected on an increasing, decreasing or constant returns to scale frontier segment and the most productive scale size a unit may attain.

DEA APPLIED

In the basic DEA model, the efficiency rating is derived as the optimal ratio of the sum of weighted outputs to inputs without any restrictions on the weights. However, in many cases the decision maker has information or judgements which can be incorporated in the model to make the assessment intuitively more valid. For example, if, when assessing policing, the outputs are types of crime solved, one may restrict the model to give a higher

weight to solving violent crimes rather than burglaries. That is, unit for unit a violent crime requires more resource than a burglary, driven by societal values. The restriction would be in the form of a relationship between the weights, rather than an exact rate of substitution. For a full explanation of how value judgements or other information can be incorporated in DEA see Thanassoulis (2001). When data over a sequence of periods of time is available (i.e. we have panel data) we can assess efficiency and productivity changes over time. However, it is important to understand that measures of efficiency across time may not be comparable. This is because DEA efficiency measures are relative to a frontier specific to a time period and that frontier may move over time. It is possible, though, to measure productivity change over time reflecting the combined effect of the change in a unit’s efficiency over time and the movement of the frontiers against which those efficiencies have been measured. Since Charnes, Cooper and Rhodes’ seminal paper, DEA has grown into one of the methods of choice for measuring and managing performance, especially so in complex contexts where multiple incommensurate inputs are used to secure multiple incommensurate outputs. It is especially useful in cases where no input or output market prices exist, yet efficient resource use is important (e.g. health, education, justice, policing etc.). DEA is used extensively across Europe and beyond in the regulation of utilities. Regulators assess the potential for efficiency savings by companies so they can factor it in their price determinations. In a survey of DEA applications covering the period 1978–2010 Liu et al. (2013) list over

3100 papers where some real life application was embedded. The five top areas in terms of application were banking, agriculture, education, health care and transportation. Software to implement many of the models available in DEA literature is widely available. For much more information on DEA including bibliography, software, and events visit www.deazone. com. Emmanuel Thanassoulis is Professor in Management Sciences at Aston Business School, Aston University, Birmingham UK. He has authored many articles, chapters and books on DEA. He has acted as consultant on efficiency and

productivity analysis to a variety of organisations, including Ofwat, Royal Mail, Severn Trent Water, and the Department for Communities and Local Government. Maria Conceição A. Silva is Associate Professor at Católica Porto Business School in Portugal. She has a PhD from Aston University, and has authored numerous articles and some book chapters on the theory and practice of efficiency and productivity measurement through DEA. She is associate editor of some international journals.

FOR FURTHER READING Charnes, A., W.W. Cooper and E. Rhodes (1978). Measuring efficiency of decision-making units. European Journal of Operational Research 2: 429–444. Liu, J.S., L.Y.Y. Lu, W-M. Lu and B.J.Y. Lin (2013). A survey of DEA applications. Omega 41: 893–902. Thanassoulis, E. (2001). Introduction to the Theory and Application of Data Envelopment Analysis. Springer.

IMPACT | SPRING 2018

TOM BONESS AND HANNAH MAYES

HEALTHSHARE NSW operates a Patient Transport Service (PTS) within the State of New South Wales (NSW), Australia, providing an important resource for patients who require transport to and from health facilities but do not need an emergency ambulance. The transport planning process aims to strike a balance between timeliness of service and efficient utilisation of resources. It is particularly challenging to

maintain this balance over a range of geographies which vary from large rural areas, such as Northern NSW Local Health District (LHD), to the urban Greater Sydney Metropolitan area. If there is insufficient capacity in the PTS system to facilitate all journeys, then some will be passed to an alternative provider for transport at an additional cost. HealthShare NSW asked consultants at Operational Research in Health

I M P R OV I N G PAT I E N T TRANSPORT IN NEW S O U T H WA L E S

Limited (ORH) to explore resourcing options for reducing the use of alternative providers in Northern NSW LHD. Through a process of analysis and modelling, ORH were able to recommend targeted additional PTS shifts to significantly reduce the cost of using these providers. Jennifer Van Cleef, Director of Patient Support Services, HealthShare NSW: ‘HealthShare NSW has been partnering with ORH since we first centralised non-emergency patient transport in 2014. Over the years, ORH has provided a significant amount of data analysis and modelling focusing on maximising efficiency, timeliness of service and fleet utilisation. This type of modelling had never been undertaken before in the scope of non-emergency patient transport and provided valuable insights as we transitioned to a new model of patient transport in NSW. The results of ORH’s modelling has allowed HealthShare NSW to build a more responsive and flexible patient transport service which has significant benefits to patients right across NSW.’

ORH has provided a significant amount of data analysis and modelling focusing on maximising efficiency, timeliness of service and fleet utilisation. This type of modelling provided valuable insights as we transitioned to a new model of patient transport in NSW

PATIENT TRANSPORT IN NEW SOUTH WALES

Historically, patient transport in NSW has been provided by a combination of

New South Wales Ambulance (NSWA), LHDs and other non-profit, voluntary and private providers. NSWA operated a dedicated non-emergency (‘green’) fleet of vehicles for providing patient transport, but the emergency (‘red’) fleet often had to cover PTS journeys, therefore making vehicles less available for emergency incidents. In 2012 the Minister for Health for NSW announced the Ambulance Reform Plan. As part of this plan, PTS was established as a separate service, combining the NSWA green fleet with the patient transport fleets run by the LHDs. A central booking hub (based in Parramatta, Sydney) went live in 2014, co-ordinating PTS in the Greater Sydney Metropolitan area. Patient transport in rural and regional LHDs in NSW is coordinated from satellites of this hub. For Northern NSW LHD, this satellite is based in Port Macquarie and went live in February 2016. The aim of this separation and centralisation was to enable more efficient and consistent use of vehicles. Since 2014 the proportion of patient transports handled by the NSWA red

fleet has dropped significantly, allowing these crews to focus on high acuity calls.

NORTHERN NEW SOUTH WALES LHD

Northern NSW LHD spans an area of 20,732km2 in northeast NSW and had a population of approximately 288,000 at the time of the 2011 Census. In 2016 the projected population was approximately 302,000 people. There are 12 public hospitals, and PTS vehicles operate from four bases: Grafton, Lismore, Maclean and Tweed Heads. In 2016/17 the PTS fleet coordinated by HealthShare NSW undertook around 200 patient transports per week in the LHD, and deployed an average of 492 weekly vehicle hours in order to meet this demand. For capacity reasons, a number of additional transports could not be handled by the PTS fleet and had to be undertaken by alternative providers (including NSWA). There is a significant additional cost to the LHD associated with transferring patients in this way. ORH was commissioned to

IMPACT | spring 2018

Patient

Journey Request

Mobility/Clinical Requirements

Infectious?

• Priority • Origin and destination

Vehicle capacity and loading restrictions

• Target arrival or pickup time • Time of call

Vehicle Type

Staff Shift

Vehicle & Crew

FIGURE 1 DEMAND AND RESOURCING

look into how PTS resourcing could be changed to reduce the number of transfers undertaken by other providers, and potentially reduce cost.

HOW ORH MODELS PTS OPERATIONS

ORH is a management consultancy that uses Operational Research (O.R.) techniques to solve resourcing and locational planning problems. ORH’s clients are principally emergency services, health authorities and sports bodies. These clients face many different challenges, but they share a need to optimise performance and deploy valuable resources in the most costeffective and efficient way. ORH uses sophisticated analysis and modelling techniques to deliver robust consultancy and software solutions that are objective, evidence-based and quantified. ORH worked with HealthShare NSW during the transition to a single PTS fleet and centralised booking hub, carrying out studies of the resources required to undertake efficient patient transport in the Greater Sydney Metropolitan Area and other health districts.

IMPACT | SPRING 2018

The ORH modelling process begins with detailed analysis of the historical journey profile, the timeliness of patient transfers against Key Performance Indicators (KPIs), and vehicle utilisation, in order to build a detailed picture of how the service is operating. The results of this analysis are then used to generate inputs for ORH’s patient transport simulation model. This simulation model replicates the key features and processes within the service. Once it has been configured and validated against the current service profile, it can be used to explore how key measures of patient timeliness and resource utilisation are affected by changes to factors such as shift patterns, vehicle type/crew mix, and the distribution and volume of patient journeys.

SIMULATING PATIENT TRANSPORT SERVICES

ORH’s patient transport simulation model is called PTSim. It is designed specifically for modelling real-world patient transport services which operate over wide geographical areas and carry

large numbers of passengers using a range of vehicle types. PTSim simulates the movements of a fleet of vehicles as they transport patients over many weeks of simulated time. Key inputs include (Figure 1): • A list of transport requests which forms the demand to simulate. Each request represents a patient needing transport to or from a medical facility and has a target time at which the transport should start or finish. • A schedule detailing the availability of staff and vehicles and the locations at which they are deployed. • Geographical information, including the locations of medical centres, pick-up and drop-off locations for each patient, and a calibrated set of travel times based on the road network. There are a number of constraints that limit how patients can be transported and how long it takes to transfer them to and from vehicles, so these are also encapsulated in the model inputs. Patients have a range of mobilities and clinical needs, so an appropriate vehicle type and crew must be selected to transport them. Specialist (multi-crewed) vehicles are provided for patients who need assistance or are carried on a stretcher. Some patients, for example, children, also need to be accompanied by an escort. Patients may also be infectious and therefore cannot travel with others. PTS providers face a number of planning challenges that must be accommodated when modelling. Many providers handle hundreds of requests per day with fleets of tens or hundreds of vehicles; this means they have to tackle large vehicle routing problems in order to plan how patients will be transported. In some cases, a significant number of transport requests are made on the

Yes

Wait at or travel to patient origin

Shift start

Off shift

Pick up patient

Pick up another patient?

Travel to selected patient destination

Yes

New route No

Waiting Shift end

Select next drop off destination

Remaining passengers?

Drop off patient

Meal break end

Meal break start

Travel to meal break location and take break FIGURE 2 MODELLED VEHICLE ACTIVITY

day of the journey instead of being pre-booked. For LHDs in NSW, typically 40%-70% of requests are booked on the same day as the required transport (approximately 60% in Northern NSW LHD). Providers therefore have to adapt their planned routes as these requests come in. The time at which each request is made is an important input to the simulation model and determines when transport can be pre-planned for that journey.

A cost function is used to compare different possible routes and decide which is ‘better’. This function takes into account patient timeliness and the total distance travelled by all vehicles. Cost parameters like these are widely used in vehicle routing algorithms. In this case, they represent the fact that PTS providers work to transport patients on time whilst ensuring resources are used efficiently.

ROUTE PLANNING IN PTSIM

This recommendation projected a significant cost saving

PTSim operates by planning efficient routes for a fleet of vehicles to carry out pre-booked transport requests (requests made before the day of transport), and then simulates the impact of changes to these plans due to certain events and interruptions such as short-notice, same-day transport requests. A fast ‘greedy’ algorithm is used to rapidly generate vehicle routes. These routes are then iteratively improved using journey insertion and removal heuristics. This is a computationally expensive process, so some assumptions are made in order to narrow down the search for routes. For example, for any given route, it is assumed a vehicle will pick up all the patients for the route before dropping them off (Figure 2). Analysis has shown that this is true in the majority of cases.

A virtual replica of the PTS network in Northern NSW LHD was created in PTSim. The model was then used to test a range of options for increasing PTS capacity. In addition to those transported by PTS, analysis showed there were 60 patient journeys per week being requested through the PTS booking satellite that were transported by an alternative provider. PTSim was used to identify the best times and locations for deploying additional PTS vehicles to transport these patients and improve timeliness. Based on this modelling work ORH made recommendations to enable 80% of the 60 patient journeys per week to

RESULTS

be undertaken with the addition of 176 weekly vehicle hours (above the existing 492 vehicle hours). It was not possible to transfer the remaining 20% without sacrificing service efficiency, as the patients were sparsely distributed across the LHD. The cost to the LHD of deploying an additional 176 PTS vehicle hours per week was far less than the cost to outsource the transports. This recommendation therefore projected a significant cost saving for the LHD. The new shifts were prioritised in terms of largest to smallest cost saving and Northern NSW began to successfully deploy the highest priority shift, operating on a Sunday. Analysis of new data has shown this vehicle to be well utilised and the use of alternative providers has reduced in line with ORH’s results. Tom Boness is a software engineer at ORH, where he specialises in designing and writing simulation models of emergency fire and ambulance services and non-emergency patient transport providers. Hannah Mayes is a consultant at ORH, a role that includes carrying out simulation and optimisation modelling for a number of ORH’s ambulance and non-emergency patient transport service clients.

IMPACT | spring 2018

MALGORITHMS Geoff Royston

Let’s look at some examples, human and computerised, of what (borrowing from Private Eye) I shall term ‘malgorithms’.

STOP AND SEARCH?

Three issues back, in a column headed ‘Life Rithms’, I wrote about how we might find it useful in our daily life to borrow some ideas from methods that computers use in algorithmic problem solving (algorithm: ‘a process or set of rules to be followed in calculations or other problem-solving operations, especially by a computer’). This time I am going to look at the darker side of algorithms – whether based on the informal models we all run in our heads or on formal ones run on computers – and at what we can do to protect ourselves and others from their misuse.

Weapons of Math Destruction looks at how mathematical models and associated algorithms are used increasingly to make decisions

This has been prompted by the book Weapons of Math Destruction by Cathy O’Neil. This book looks at how mathematical models and associated algorithms are used increasingly to make decisions, for example in the justice system, in education, and in the workplace, and how, when such models and algorithms are opaque, unregulated, or just wrong, they can damage lives, increase inequality and even threaten democracy. (The book’s title relates to the ability of algorithms, when automated, to be deployed at huge scale and so for harmful ones to wreak wide-ranging damage.)

Suppose there has been an outbreak of burglaries in my locality and I am the police officer on the beat. I see someone with green hair walking toward me. Should I stop and search him (or her) for evidence of burgling? If, round here, green-haired people are more likely than others to have convictions for burglary, then this might be seen as a reason for having rules - an algorithm - for stop and search that lead to a focus on that group (there may of course be other motivators, including pure pileous prejudice). But first look at some figures. Suppose this is a crime-ridden locality where about 5% of the population are burglars. 1% of the people here have green hair and historically members of this group have been found to be twice as likely as average to be burglars. Then, out of a thousand people in the area, their expected composition would be: Green-haired

Non-green-haired

Burglar

Non-burglar

940

Notwithstanding the greater criminality of the local green-hairs, they are committing only 2% of the burglaries in this area. Further, although the chance that a stop and search of a green-haired person will detect a burglar is twice the chance of that for a search of anyone else (assuming for simplicity that burglars always carry evidence of their trade with them and that searches never make mistakes!), 90% of them will be found to be not carrying anything incriminating. A broad brush can tar a lot of innocent people. Some elementary quantitative analysis adds further insight. Consider two possible stop and search approaches (a) focus on the higher risk group (b) pick a representative sample. Suppose there are enough resources to stop and search 10% of a crowd of 10,000 people. Then, using the figures for the above example, we have:

Focus on the higher

Number of green-

Number of non-

haired people

green-haired people

searched

detected by search

green haired people

100

900

10/55 = 18%

990

1/51 = 2%

Number of burglars % of arrests that are of

risk group Pick a representative sample

As so often, there is a trade off between efficiency and equity. Here the strategy ‘focus on the higher risk group’ slightly increases the efficiency of the search process in terms of arrests per search (though still without much impact on the overall numbers of burglars remaining at large) but vastly increases the proportion of arrests that are of green-haired people. This results in them being hugely over-represented in arrests compared to that group’s contribution to crime. Adoption of a strategy of this broad brush kind would suggest that equity was being given an extremely low weighting compared to efficiency; which would certainly not do anything for relations with the green-haired community. For a stop and search rule this looks like a ‘malgorithm’. There is also a more subtle effect of this kind of selective approach, even for one less extreme than that shown above: the creation of pernicious feedback loops. Disproportionate representation of a particular group in arrests and convictions, could tend to reinforce any prejudice about the group having higher crime rates; which could lead to them being subjected to even more stop and search. (Parliament has taken steps to mitigate the problem of police ‘broad brush’ stop and search - the law now says that - apart from some very special exceptions - there must be reasonable, specific, suspicion that someone is carrying a prohibited item, and that an officer must not base their suspicion on generalities such as race, colour, age or the way someone dresses.)

HIRE AND FIRE?

As Cathy O’Neil points out in her book, ‘our livelihoods increasingly depend on our ability to make our case to machines’. For example, for many businesses, success requires appearing near the top of a Google search. And something similar can apply for individuals seeking to be granted a loan or chosen for a job; applications are often sifted initially by an automated process.

This can go badly wrong; Weapons of Math Destruction discusses the infamous case of St George’s Hospital Medical School in South London back in the 1980s. Seeking to improve the efficiency of processing large numbers of applications from prospective students, St George’s had designed a computerised sifting algorithm, using past sift data as a learning set to emulate human selection procedures. Unfortunately, the computer learnt to emulate these all too well. It learnt that past sifts had rejected applications written in poor English, and that these correlated with applicants with foreign birthplaces. It learnt that female applicants were often rejected. The computerised selection duly followed suit. Few staff understood its internal workings. Eventually – after some years - somebody checked. In 1988 the medical school was found guilty of racial and gender discrimination in its admissions policy. St George’s may have been an early adopter of a weapon of math destruction for candidate selection, but it certainly was not the last. For example, LinkedIn Recruiter has been criticised for its ‘The People You May Want to Hire’ feature. This used members’ public data and suggested relevant candidates. Race and gender were not explicitly included but other characteristics like location were, and could over time increasingly influence the algorithm’s selection preferences. Another target at which Weapons of Math Destruction aims some fire are scheduling algorithms. It notes that in the ‘gig’ economy, computerised scheduling systems, seeking efficiency, can tightly limit hours, change shifts at short notice or demand working at unsocial times. This can play havoc with people’s daily lives. Worse, in the longer term, damaging feedback loops can kick in when, for example, such scheduling prevents workers from fitting in further education, which then locks them into low wage jobs. Here, as often elsewhere, the main problem lies not with algorithmic software as such but with the objectives that are chosen to drive it.

IMPACT | spring 2018

Healthy algorithms are based on models that are transparent, use data that are directly related to the issue in question, and are regularly checked against the real world to learn from their errors

What about the highly sensitive workplace decision to dismiss someone or make them redundant? Cathy O’Neil cites a case where a San Francisco company devised computer software that analysed corporate e-mails to identify staff who were good at generating ideas or spreading information. Useful perhaps for, say, identifying possible collaborators. Not so good for one thing it was used for – identifying people to lay off. People flagged up by the computer system as poor generators of ideas or as weak connectors were deemed prime candidates for ‘letting go’. But the software did not include other important

IMPACT | spring 2018

performance factors, nor could its verdict be checked adequately, given that no error-correcting feedback would be available from those who had been fired.

AVOIDING MALGORITHMIC MALAISE

How can we spot and avoid malgorithms? Healthy algorithms are based on models that are transparent, use data that are directly related to the issue in question, and are regularly checked against the real world to learn from their errors. The models underlying malgorithms by contrast tend to be opaque ‘black boxes’, often use data that are poor proxies for directly relevant information, and frequently involve pernicious feedback loops that lead to them justifying their own assumptions.

Mathematical models should be our tools not our masters

On the day I write this I read a newspaper article about a traveller to the USA who had his debit card blocked by a UK bank because its automated fraud detection algorithms ‘identified a pattern of spending similar to those it has seen in cases of fraud’. A simple check by a human (or probably even a computer algorithm that focused on specifics not generalities) would have immediately shown that there was no problem. Which leads me to end with another quote from Weapons of Math Destruction: ‘Mathematical models should be our tools not our masters’. Dr Geoff Royston is a former president of the O.R. Society and a former chair of the UK Government Operational Research Service. He was head of strategic analysis and operational research in the Department of Health for England, where for almost two decades he was the professional lead for a large group of health analysts.

HEALTH SYSTEMS Apparatus bellis corrumperet Medusa, quod fiducias amputat verecundus suis. Perspicax agricolae suffragarit Augustus. Suis vocificat fiducias. Saburre miscere Aquae Sulis. Pessimus tremulus matrimonii insectat Octavius. Satis saetosus ossifragi agnascor incredibiliter perspicax apparatus bellis. Satis quinquennalis fiducias imputat gulosus agricolae. Apparatus bellis corrumperet Medusa, quod fiducias amputat verecundus suis. Apparatus bellis corrumperet Medusa, quod fiducias amputat verecundus suis. Perspicax agricolae suffragarit Augustus. Suis vocificat fiducias. Saburre miscere Aquae Sulis. Pessimus tremulus matrimonii insectat Octavius. Satis saetosus ossifragi agnascor incredibiliter perspicax apparatus bellis. Satis quinquennalis fiducias imputat gulosus agricolae. Apparatus bellis corrumperet Medusa, quod fiducias amputat verecundus suis.

JOURNAL OF THE OPERATIONAL RESEARCH SOCIETY

Contents

Health Systems is an interdisciplinary journal promoting the idea that all aspects of health and healthcare delivery can be viewed from a systems perspective. The underpinning philosophy of the journal is that health and healthcare systems are characterized by complexity and interconnectedness, VOLUME 00 NUMBER 00 MONTH 00 ISSN:else”. 0960-085X where “everything affects everything Thus, problems in healthcare need to be viewed holistically 00 as an integrated system of multiple components 00 (people, organizations, technology and resources) 00 perspectives. The journal sees the systems and approach to be widely applicable to all areas of health 00 and healthcare delivery (e.g., public health, hospitals, 00 primary care, telemedicine, disparities, community health). Hence, the principal aim of the journal is to 00 bring together critical disciplines that have proved themselves already in health, and to leverage these 00 contributions by providing a forum that brings 00 together diverse viewpoints and research approaches (qualitative, quantitative, and conceptual). 00 00

Co-editors 00 Sally Brailsford, 00 University of Southampton, UK

THE EUROP JOURNAL O INFORMATIO SYSTEMS

Paul Harper, Cardiff University, UK Nelson King, Khalifa University, United Arab Emirates Cynthia LeRouge, Florida International University, USA

VOLUME 00

T&F STEM @tandfSTEM

Dov Te’eni @tandfengineering NUMBER 00

Explore more today… bit.ly/2BXpklD MONTH 2018

Annual Analytics Summit 2018 Tuesday 12 June IET London

As part of London Technology Week, the Annual Analytics Summit delivers a one-day learning and networking event about how big data and analytics are shaping organisational decision-making. Filled with case studies, innovations and strategies on turning data into decisions, the Annual Analytics Summit is the event for practitioners and decision-makers alike.

Some of the talks and workshops scheduled include: Richard Bradley, TfN Head of Data, Analysis and Appraisal How Transport for the North are planning to transform the North’s economy through innovative approaches to data and modelling

Keith Still, Professor of Crowd Science at Manchester Metropolitan University Marcel Altenburg, Enterprise Delivery Fellow at Manchester Metropolitan University Understanding crowd dynamics

Ian Randolph, Product Owner, Data Science at Trainline Ethics for weapons of mass persuasion

Helena Guerra, Senior Manager, Big Data Technology at Vodafone Building a big data environment

Early bird rate available until 30 April 2018 £125 + VAT From 1 May £175 + VAT Price includes access to event, refreshments and buffet lunch.

#AS18 www.analytics-events.co.uk