Asleep & Awake in the ICU by Laurens Reinke

Paranimfen Dr. F. Doesburg Dr. L. Zwakman-Hessels

Cover design by Roy Monteiro Printed by Optima Grafische Communicatie Copyright © 2021 Laurens Reinke All rights reserved. No parts of this thesis may be reproduced, stored in a retrieval system, or transmitted in any form or by any means, without prior written permission of the author, or when appropriate, of the publishers of the publications included in this thesis.

Asleep and awake in the ICU

Proefschrift

ter verkrijging van de graad van doctor aan de Rijksuniversiteit Groningen op gezag van de rector magnificus prof. dr. C. Wijmenga en volgens besluit van het College voor Promoties. De openbare verdediging zal plaatsvinden op Dinsdag 1 maart 2022 om 16:15 uur

door

Laurens Reinke geboren op 16 november 1987 te Anloo

Promotores Prof. dr. J.E. Tulleken Prof. dr. A.R. Absalom Beoordelingscommissie Prof. dr. J.C.C. van der Horst Prof. dr. R.A. Hut Prof. dr. ir. N.M. Maurits

Voor Lis

Table of Contents Chapter 1

General Introduction Thesis outline

8 16

I. The measurement of sleep in the ICU Chapter 2

The clinical measurement and analysis of sleep

Chapter 3

Intensive care unit depth of sleep: Proof of concept of a simple electroencephalography index in the nonsedated

Chapter 4

Automated versus manual scoring of ICU sleep data

II. Asleep in the ICU Chapter 5

Systematic review of the effects of intensive-care-unit noise on sleep of healthy subjects and the critically ill

Chapter 6

The importance of the intensive care unit environment in sleep: A study with healthy participants

104

Chapter 7

Norepinephrine administration is associated with increased melatonin levels and daytime sleeping in critically ill patients: A retrospective observational study

126

III. Awake in the ICU Chapter 8

The effect of chronotype on sleepiness, fatigue, and psychomotor vigilance of ICU nurses during the night shift

152

Chapter 9

Sleep deprived and unprepared

174

Chapter 10

Summary and conclusions General discussion and future perspectives

180 183

Chapter 11

Samenvatting Curriculum Vitae Dankwoord

188 192 193

Chapter 1 General introduction Thesis outline

Chapter 1

General introduction Sleep Sleep is one of the most pleasant, essential and ubiquitous activities in humans. It is a dynamic, complex physiological process essential for homeostasis, recovery, and survival [1, 2]. Sleep is most commonly defined as a set of physiological characteristics observed in mammals, although even non-mammalian species such as birds, fish, and even worms exhibit sleep-like resting states [3]:     

Behavioural quiescence Altered consciousness Elevated arousal thresholds Homeostatic regulation (sleep rebound, REM rebound) Naturally and rapidly reversible

Sleep and the transitions between sleep stages are tightly controlled by complex interactions between mutually inhibiting sleep promoting and arousal systems in the brainstem, hypothalamus and basal forebrain [4]. Each night during sleep, they collectively orchestrate a myriad of physiological processes throughout the body.

Sleep architecture Based on surface measurements of brain, eye, and muscle electrical activity, sleep can be divided into two distinct phases that alternate during a night’s sleep: rapid eye movement sleep (REM) and non-REM sleep (NREM) [5]. REM is characterized by low muscle tone, rapid desynchronized low-voltage electroencephalographic (EEG) activity, and the eponymous rapid eye movements. During REM sleep, brain electrophysiological [3] and metabolic [6] activity most resembles the waking state, and dreams mostly occur during this phase of the sleep cycle. During wakefulness and REM sleep the functional connectivity within the brain is high, resulting in desynchronized brain activity as measurable using EEG [7]. During REM sleep cerebral oxygen and glucose consumption regularly exceeds that during quiet waking [6]. It is often referred to as paradoxical sleep due to these similarities to wakefulness and dissimilarity to the other stages of sleep. REM sleep occupies up to 25% of the total sleep period. It is normally preceded by a short period of NREM sleep, and the duration of REM periods may increase towards the end of the sleep period. REM sleep has a particularly high sensitivity to disruption. Particularly environmental noise increases REM sleep latency, and reduces total REM activity and REM bout duration in both experimental and practical settings [8, 9]. Non-REM (NREM) sleep is characterized and defined by the presence of specific transient waveforms, such as sleep spindles and K complexes, and a gradually deepening loss of consciousness. Despite this gradual transition, NREM is commonly divided into three discrete stages, N1, N2, and N3, with increasing thresholds for arousal, and decreasing degree of awareness. Stage N1 is thought of as a transitional stage between wakefulness and sleep, and typically accounts for 2-5% of the total sleep period. During N1, the sleeper

General introduction

is still aware of his surroundings. Whether stage N1 is functionally similar to the other NREM sleep stages is debatable, and its contribution to the overall quality of a period of sleep is unknown. The EEG during N1 resembles REM sleep, albeit without the low muscle tone and eye movements. The onset of stage N2 is defined by K-complexes and sleep spindles, and around 50% of the total sleep is occupied by stage N2 sleep. Sleep spindles are 11-15 Hz EEG oscillations that last up to 3 seconds. They are positively associated with memory performance, specifically of declarative memory [10]. Kcomplexes are very large transient EEG waveforms that occur spontaneously during N2, but may also be generated following loud noises or somatosensory or visual stimuli [11]. These waves are hypothesized to suppress unwanted cortical activation by external stimuli during sleep. During N2 the sleeper is no longer aware of his surroundings, thereby meeting most of the defining characteristics of sleep. The deepest stage of NREM sleep, N3, is also called slow-wave-sleep (SWS) as slow, high amplitude waves dominate during this stage. SWS accounts for up to 20% of the total sleep period. Prolonged wakefulness leads to increased slow-wave activity during NREM sleep. As a night of sleep progresses, the intensity of slow-waves decreases. SWS used to be divided into two separate stages of sleep, N3 and N4, where N4 was defined by an even larger abundance of slow waves than N3 [12]. This division was later abandoned in favour of a more comprehensive and practical definition. The brain transitions through these sleep stages sequentially, starting with N1 following which the depth of sleep gradually deepens to N3 before transitioning back to N2, followed by a period of REM sleep. After some time, the brain again transitions through N2 to N3 and back up to REM sleep. This cycle repeats until the sleeper is awoken. In the absence of sleep-disrupting stimuli, this means that this 90-minute cycle is repeated four to five times a night.

Chapter 1

Circadian timekeeping Sleep onset and offset are regulated by two competing processes. On one hand, sleep is promoted by the sleep homeostat, mainly driven by adenosine (process S). On the other hand, daily rhythms of wakefulness are regulated by the circadian pacemaker (process C), located in the suprachiasmatic nucleus (SCN). Most circadian rhythms are maintained by communicating the rhythm dictated by the central pacemaker to a broad range of target organs with their own pacemaker through diurnal secretion cycles of the neurohormone melatonin from the pineal gland (Figure 1, middle trace). The synthesis of melatonin from serotonin is inhibited by short-wavelength light, and is limited by hydroxylation of the essential amino-acid tryptophan. This light-dependent release of melatonin is actualized by intrinsically photosensitive retinal ganglion cells (ipRGCs), which project to the pineal gland via the circadian pacemaker. These cells with limited contribution to conscious sight contain the photopigment melanopsin in contrast to the rods and cones of the outer retina. This means that even rodless and coneless blind animals are able to maintain circadian timekeeping [13]. The central pacemaker has an intrinsic phase of just over 24 hours, and is normally entrained by a range of environmental cues, commonly known as Zeitgebers. The main factor that entrains this pacemaker is the presence and absence of short wavelength light (446-483 nm, Figure 1, bottom trace) during the day and the night, respectively [14, 15]. Other exogenous cues that entrain the human circadian pacemaker are temperature, eating patterns, physical and cognitive exercise, and social interaction. The analysis of circadian rhythm in humans relies on the measurement of a limited set of phase markers for circadian processes. These include core body temperature, melatonin or cortisol concentration in plasma, and the urinary concentration of the melatonin metabolite 6-sulfatoxymelatonin. In healthy subjects and under controlled circumstances these markers provide robust phase estimations for the central pacemaker. Fever, organ dysfunction, inflammation, and certain medications may however alter or mask these markers, rendering these measurements unreliable [16, 17].

General introduction

Sleep distribution Wake REM N1 N2 SWS MT

Melatonin concentration (pmol/L)

18:00 21:00 00:00 03:00 06:00 09:00 12:00 15:00 18:00 21:00 00:00 03:00 06:00 09:00 12:00 15:00

Melatonin 400

Peak 03:51

Onset 23:18

200

Onset 23:22

100 0

18:00 21:00 00:00 03:00 06:00 09:00 12:00 15:00 18:00 21:00 00:00 03:00 06:00 09:00 12:00 15:00

Light

10 Light intensity (lux)

Peak 04:03

300

DAY

NIGHT

DAY

NIGHT

DAY

−1

−2

18:00 21:00 00:00 03:00 06:00 09:00 12:00 15:00 18:00 21:00 00:00 03:00 06:00 09:00 12:00 15:00 Time of day (hours:minutes)

Figure 1. Intensive care unit (ICU) example of the inhibition of melatonin release (middle trace) by environmental light (bottom trace), and the effect on the distribution of sleep (top trace). Around 06:00 natural light starts to slowly illuminate the ICU until artificial light sources are turned on causing relatively stable light intensity for most of the day until artificial light sources are shut off at around 22:30 (bottom trace). Once environmental light intensity goes below a certain threshold, blood melatonin concentrations start to rise, before peaking around 04:00 (middle trace, interpolated using splines). Once environmental light intensity exceeds a certain threshold blood melatonin concentration decreases and remains low for the remainder of the day. Sleep (top trace) is largely confined to periods with sufficient sleep pressure (not shown), low light, and high melatonin

Chapter 1

The function of sleep During sleep, animals have limited defences against predators, and sacrifice valuable time that otherwise would have been spent searching for food or mating. The evolutionary benefit of this trade-off must outweigh the sacrifice but continues to be a subject of scientific debate and investigation. The functional importance of sleep is perhaps best illustrated by observing the detrimental effects of sleep deprivation. Continued complete sleep deprivation has long been known to be lethal [18], but even mild deprivation has measurable effects on human behaviour, cognition, and wellbeing that have yet to be fully understood. Although it is difficult to deprive study subjects from REM sleep without impacting NREM sleep, the consensus is that it plays a central role in the formation and consolidation of certain types of memory [19]. The unique muscle activity associated with REM sleep was initially thought to be a by-product, but new evidence shows that they are likely triggered against a background of atonia to facilitate targeted development of the sensorimotor system [20]. This concept is further supported by the observation that the amount of REM sleep decreases with age [21, 22]. During NREM sleep, brain metabolism is greatly reduced, although the old concept of sleep as a means of energy conservation is not supported by most experimental evidence [3]. A more recent study suggests that the interstitial space of the mammalian brain increases during sleep, enhancing the clearance of extracellular wastes and toxins from the cerebrospinal fluid (CSF) [23]. Interestingly, the degree of clearance decreases when body posture during sleep is limited to less favourable positions [24]. Additionally, the deepest stages of sleep are particularly important in declarative memory consolidation [25]. Collectively, these data support the hypothesis that sleep evolved to utilize less valuable time to recover and tidy up the stressed brain through increased waste clearance to maintain homeostasis and to optimize and consolidate memory through targeted feedback.

General introduction

Sleep disruption and intensive care Disrupted or delayed sleep is associated with impaired immune function [26], increased susceptibility to infections and impaired wound healing [27, 28], impaired carbohydrate metabolism and endocrine function [29], increased pain perception [30, 31], and impairment of neurophysiologic organization and memory consolidation [32]. When a recent Dutch nation-wide survey investigated the quantity and quality of sleep of patients receiving regular care, they found that the total sleep time was reduced by 83 minutes during admission compared to sleep at home [33]. More than half of this difference could be attributed to earlier final awakening, but patients also reported experiencing more awakenings throughout a night in the hospital. As in this study, patients often report noise from other patients and hospital staff to be most distracting from sleep. Out of approximately 1.7 million overnight hospital admissions in the Netherlands in 2018, 76.103 admissions were to an intensive care unit (ICU) to receive intensive care [34]. The median length of ICU stay was 1.1 days, with 25% of patients staying 2.8 days or longer [34]. During this period of intensive care, many patients experience even more difficulty sleeping, and sleeping well. Sleep deprivation is common among hospital patients, particularly in the ICU [35], where it affects up to 60% of all admitted patients [36]. Among ICU patients, sleep is often fragmented by frequent arousals and awakenings. For ICU patients, half of the total sleep time occurs during the day [37, 38], and total daily sleep duration is reduced [28, 36, 39]. In intensive care unit (ICU) patients prolonged sleep disruption may lead to the development of delirium, prolonged admission and increased mortality risk [2, 32, 40] in addition to the aforementioned risks associated with hospital admission and illness. Sleep disruption and delirium are widely recognized as major care complications that disrupt the workflow in the ICU. Consequently, an increasing number of ICU guidelines mention the importance of sleep in some form as a relevant aspect of patient management and ICU design [41–44]. These guidelines commonly recommend avoiding nightly bed-side interventions, limiting environmental noise, and tailoring the scheduling of medication and nursing care rounds to avoid further sleep disruption or delirium. The concerns about the potentially detrimental clinical effects of ICU sleep disruption have led to a sharp rise in interest for the topic of sleep in the ICU, with more than half of the available publications on the subject being published in the last five years. The ICU is a unique environment where a multitude of intrinsic and extrinsic factors may hamper sleep [45–51]. The aetiology of sleep disturbance among individual ICU patients then is likely to be multi-factorial and different between patients. Noise pollution, stress, pain and critical illness, may increase arousal and decrease the activity of sleep promoting systems. Sleep disruption and delirium are both commonly observed concurrently with altered patterns of melatonin secretion [35, 52]. More specifically, the observed fragmentation and abnormal distribution of sleep is often associated with an impaired melatonin biorhythm [53–58]. When melatonin secretion patterns are

Chapter 1

insufficiently synchronized with external rhythms, the biological clock is said to be ‘running free’ (see Figure 2). Over extended periods of time this unwound body clock may result in activation of arousal systems during the night and the homeostatically driven promotion of sleep during the day. This tendency to sleep during the day, when the ICU environment is even less conducive to sleep, is often observed and reported [28, 37, 59, 60].

Normal contrast between Normalday contrast andbetween night day and night

No contrast between No contrast between day and day and night night

Zeitgebers

Zeitgebers Light (short wavelength) Light (short wavelength)

Social interaction Social interaction

Eating/drinking Eating/drinking

Exercise Exercise

+ + + +

-|| Entrainment Entrainment ↓↓

Day 1

Sleep

Day 2

Sleep

Day 3 Day 4

Day 1 Day 2 Day 3 Day 4

Sleep Sleep Sleep Sleep Time of day → Sleep Time of day → Sleep

++ ++ ++ ++

+/+/+/+/-

+/-

+/- +/-

+/- +/- +/+/- +/- +/- +/+/- +/- +/- +/| | Free-running Free-running ↓ ↓ +/-

+/+/+/+/-

Sleep

Sleep Sleep Sleep Sleep Sleep Sleep Time of day → Sleep Time of day →

Figure 2. Entrainment and free-running of the circadian pacemaker depending on the presence and absence of diurnal variation of exposure to Zeitgebers. Normally the amount of short wavelength light and other Zeitgebers fluctuates over each 24-hour period (top left). The clear contrast between day and night entrains the circadian pacemaker and promotes naturally occurring sleep during the night (lower left). Once admitted to the ICU, patients experience much less contrast in the exposure to light and other Zeitgebers (due to intensive care, immobility, enteral tube feeding, etc.) effectively blurring the line between day and night (top right). Theoretically, this can cause free-running of the circadian pacemaker, and circadian sleep rhythms are lost (lower right)

General introduction

Improving sleep in the ICU When an international group of over 500 hospitals in Europe and Canada asked their ICU nurses about practices around patients’ sleep, they found that only 18% of the ICUs were using non-pharmacological interventions, 59% used benzodiazepines, and 18% used melatonin supplements to promote sleep. The most commonly reported interventions were staff noise reduction, limiting nocturnal artificial light and nursing interventions, and keeping patients awake during the day. Although 72% of nurses would like to implement protocols to improve sleep, only 9% had, and only 1% used tools to assess subjective sleep. Arousal limiting interventions such as behavioural noise reduction protocols, sound masking, earplugs, and eye-masks at night are affordable and relatively well-examined methods to improve sleep in the ICU [9, 46–48, 50, 61–63]. These studies show varying results ranging from small reductions in arousal causing factors without significant effects on sleep, to significant improvements of perceived sleep quality [64]. Recently, interventions with dynamic lighting have been attempted to decrease the incidence of delirium and improve the distribution of sleep by entrainment of the biological clock, but did not decrease the incidence of delirium [65], or improve melatonin secretion [55]. Nocturnal enteral supplementation of melatonin has been shown to reduce the need for sedation among ICU patients [66], but does not generally lead to other clinically significant improvements in sleep. It is important to note that the degree of sleep disruption among individual patients in these investigated samples varies widely, although it remains unclear what causes the sleep of some ICU patients to be much more affected. The causes of non-response to interventions in the physiology, treatment, and environment of most ICU patients similarly remain unclear. This is often attributed to the heterogeneity of the population and relatively small sample sizes of objective sleep measurements. Sleep promoting interventions such as earplugs [9, 61], eye-masks at night [9], and dynamic lighting [55, 67, 68] are very cost-effective, particularly when compared with the use of antipsychotic medications [69]. The causes of non-response to these interventions observed in some patients remains unclear [9, 61]. Identification of the exact individual causes would enable targeted and even more cost-effective application of these interventions. Although most individual sleep-disrupting stimuli are theoretically modifiable, we know very little about relative importance, interaction with other relevant aspects of critical illness and treatment. More fundamentally, it would likely require continuous, reliable, and scalable multimodal monitoring to tailor treatment and the immediate environment to an individual ICU patient’s needs. Furthermore, any changes made to this immediate environment are likely to affect medical and nursing staff as well.

Chapter 1

Thesis outline This thesis takes a holistic approach to the study and improvement of ICU sleep by first (I.) investigating existing and new approaches to the measurement and scoring of ICU sleep, (II.) questioning previously made assumptions from scientific literature about ICU sleep disruption and proposing new avenues for further research, and (III.) discussing the complexities that spring from the conflicting interests between ICU staff and critically ill patients being in the same environment at night. Combined, they form the first step towards a more diligent and productive approach to improving sleep in the ICU.

I. The measurement of sleep in the ICU In chapter 2 we will establish a comprehensive overview of tools and techniques available for sleep analysis. A simple scalable tool to get quick and real-time insight into the sleep depth of ICU patients would be helpful in daily practice and, in theory, would offer therapeutic possibilities for optimal timing of care procedures and of exposure to environmental cues to entrain circadian rhythms, but does not currently exist. We will define the requirements for such a scalable, reliable, and near-real-time system and discuss important considerations for its application in the ICU. Chapter 3 will describe a proof of concept for a system theoretically capable of near-realtime EEG-based sleep depth analysis, and discuss its face validity in a small sample of non-sedated ICU patients. To gain more insight into the pathophysiological processes and aetiology of sleep disturbances during ICU stay and during severe illness, the more extensive and expensive method of polysomnography (PSG) with R&K scoring by human expert is still considered the gold standard. PSG is of proven value in non-ICU patients, but a significant drawback is that scoring of the data requires a lot of time and specific expertise which significantly limits scalability. An automated scoring system is on hand but is not yet validated for the ICU patient population. In chapter 4 we will therefore compare the manual scoring of a larger sample of ICU polysomnographic recordings by a human expert to automated scoring by a commercially available system to inform future attempts to better understand and improve ICU patients’ sleep.

II. Asleep in the ICU Previous studies investigating the aetiology of ICU sleep disruption have often attributed most of the observed disruptions to excessive environmental noise and a general lack of external cues for robust internal timekeeping. As a result, efforts to improve ICU sleep have been aimed almost exclusively at minimizing noise and other modifiable environmental factors. In chapter 5 we will systematically review experimental studies that aim to reduce the effects of environmental noise on sleep, and attempt to perform meta-analysis of their results to determine the relative importance of excessive noise in disrupting ICU sleep.

Thesis outline

As the full impact of the ICU environment on patients’ sleep is almost impossible to isolate from disease and treatment-related factors, we will describe the cumulative impact of the ICU environment on the objective and subjective quality of sleep experienced by healthy subjects in chapter 6. As previously discussed, little is known about the complex interaction between biorhythms and the circadian distribution of sleep in ICU patients. In chapter 7 we will study the effect of certain commonly used ICU medications on circadian timekeeping and the distribution of sleep. By simultaneous observation of pharmacological treatments, melatonin rhythms, and sleep we may expose factors that inadvertently disturb regular circadian timekeeping and sleep in unforeseen ways.

III. Awake in the ICU In anticipation of future technological, logistical and behavioural optimizations of intensive care environments and intensive care processes to better facilitate sleep for ICU patients, we will investigate how ICU staff cope with the demand for intensive care around the clock, in this shared environment. In chapter 8 we will study the existing strategies that ICU nurses apply to stay vigilant and productive under pressure of acute sleep deprivation. The impact of unique individual characteristics and demographics on relevant estimates of productivity and psychomotor vigilance during night shifts will be determined to inform recommendations for optimal intensive care environments where both staff and patient are considered. Chapter 9 will provide a further discussion of ways to improve the resilience of night shift teams through non-technical means.

Chapter 1

References 1. 2.

4. 5.

10.

11. 12.

13.

14.

15. 16.

Weinhouse GL, Schwab RJ. Sleep in the critically ill patient. Sleep. 2006;29:707–716. Kamdar BB, Needham DM, Collop NA. Sleep Deprivation in Critical Illness: Its Role in Physical and Psychological Recovery. J Intensive Care Med. 2012;27:97–111. Allada R, Siegel JM. Unearthing the Phylogenetic Roots of Sleep. Curr Biol. 2008;18:R670–R679. Saper CB, Fuller PM, Pedersen NP, et al. Sleep State Switching. Neuron. 2010;68:1023–1042. Aserinsky E, Kleitman N. Regularly occurring periods of eye motility, and concomitant phenomena, during sleep. Science. 1953;118:273–4. Madsen PL, Schmidt JF, Wildschiodtz G, et al. Cerebral O2 metabolism and cerebral blood flow in humans during deep and rapid-eyemovement sleep. J Appl Physiol. 1991;70:2597– 2601. Massimini M. Breakdown of Cortical Effective Connectivity During Sleep. Science. 2005;309:2228–2232. Topf M, Davis JE. Critical care unit noise and rapid eye movement (REM) sleep. Heart Lung. 1993;22:252–8. Hu RF, Jiang XY, Zeng YM, et al. Effects of earplugs and eye masks on nocturnal sleep, melatonin and cortisol in a simulated intensive care unit environment. Crit Care. 2010;14:R66. Tamminen J, Payne JD, Stickgold R, et al. Sleep spindle activity is associated with the integration of new memories and existing knowledge. J Neurosci. 2010;30:14356–60. Colrain IM. The K-complex: A 7-decade history. Sleep. 2005;28:255–273. Kales A, Rechtschaffen A. A Manual of Standardized Terminology, Techniques and Scoring System for Sleep Stages of Human Subjects. U.S. National Institute of Neurological Diseases and Blindness, Neurological Information Network, Bethesda, Md. Provencio I, Rodriguez IR, Jiang G, et al. A Novel Human Opsin in the Inner Retina. 2000;20:600– 605. Brainard GC, Hanifin JP, Greeson JM, et al. Action spectrum for melatonin regulation in humans: evidence for a novel circadian photoreceptor. J Neurosci. 2001;21:6405–12. Chan M-C, Spieth PM, Quinn K, et al. Circadian rhythms. Crit Care Med. 2012;40:246–253. Gehlbach BK, Chapotot F, Leproult R, et al. Temporal disorganization of circadian rhythmicity and sleep-wake regulation in mechanically ventilated patients receiving

17.

18.

19.

20.

21.

22.

23.

24.

25.

26.

27.

28.

29.

30.

31.

continuous intravenous sedation. Sleep. 2012;35:1105–1114. Gazendam JAC, Dongen HPA Van, Grant DA, et al. Altered Circadian Rhythmicity in Patients in the ICU. 2013;483–489. Manasseina MM. Quelques observations expérimentales sur l’influence de l’insomnie absolue. Arch Ital Biol. 1894;322–5. Peever J, Fuller PM. Neuroscience: A Distributed Neural Network Controls REM Sleep. Curr Biol. 2016;26:R34–R35. Sokoloff G, Uitermarkt BD, Blumberg MS. REM sleep twitches rouse nascent cerebellar circuits: Implications for sensorimotor development. Dev Neurobiol. 2015;75:1140–1153. Reynolds CF, Kupfer DJ, Taska LS, et al. EEG sleep in elderly depressed, demented, and healthy subjects. Biol Psychiatry. 1985;20:431– 442. Lancee J, Eisma MC, van Straten A, Kamphuis JH. Sleep-Related Safety Behaviors and Dysfunctional Beliefs Mediate the Efficacy of Online CBT for Insomnia: A Randomized Controlled Trial. Cogn Behav Ther. 2015;44:406– 422. Xie L, Kang H, Xu Q, et al. Sleep drives metabolite clearance from the adult brain. Science. 2013;342:373–377. Lee H, Xie L, Yu M, et al. The Effect of Body Posture on Brain Glymphatic Transport. J Neurosci. 2015;35:11034–11044. Marshall L, Helgadóttir H, Mölle M, Born J. Boosting slow oscillations during sleep potentiates memory. Nature. 2006;30;444(7119):610-3. Irwin M, McClintick J, Costlow C, et al. Partial night sleep deprivation reduces natural killer and celhdar immune responses in humans. FASEB J. 1996;10:643–653. Friese RS, Diaz-Arrastia R, McBride D, et al. Quantity and quality of sleep in the surgical intensive care unit: are our patients sleeping? J Trauma. 2007;63:1210–1214. Cooper AB, Thornley KS, Young GB, et al. Sleep in critically ill patients requiring mechanical ventilation. Chest. 2000;117:809–818. Spiegel K, Leproult R, Van Cauter E. Impact of sleep debt on metabolic and endocrine function. Lancet. 1999;354:1435–1439. Lautenbacher S, Kundermann B, Krieg JC. Sleep deprivation and pain perception. Sleep Med Rev. 2006;10:357–369. Onen SH, Alloui A, Gross A, et al. The effects of total sleep deprivation, selective sleep interruption and sleep recovery on pain

General introduction

32.

33.

34. 35.

36.

37.

38.

39.

40.

41.

42.

43.

44.

45.

46.

tolerance thresholds in healthy subjects. J Sleep Res. 2001;10:35–42. Boyko Y, Ørding H, Jennum P. Sleep disturbances in critically ill patients in ICU: How much do we know? Acta Anaesthesiol Scand. 2012;56:950–958. Wesselius HM, van den Ende ES, Alsma J, et al. Quality and Quantity of Sleep and Factors Associated With Sleep Disturbance in Hospitalized Patients. 2020;178:1201–1208. Nationale Intensive Care Evaluatie. Jaarboek 2018 Stichting NICE. Mistraletti G, Carloni E, Cigada M, et al. Sleep and delirium in the intensive care unit. Minerva Anestesiol. 2008;74:329–33. Freedman NS, Kotzer N, Schwab RJ. Patient perception of sleep quality and etiology of sleep disruption in the intensive care unit. Am J Respir Crit Care Med. 1999;159:1155–1162. Bourne RS, Minelli C, Mills GH, Kandler R. Clinical review: Sleep measurement in critical care patients: research and clinical implications. Crit Care. 2007;11:226. Parthasarathy S, Tobin MJ. Applied Physiology in Intensive Care Medicine. Springer Berlin Heidelberg, Berlin, Heidelberg. Friese RS, Diaz-Arrastia R, McBride D, et al. Quantity and quality of sleep in the surgical intensive care unit: are our patients sleeping? J Trauma. 2007;63:1210–4. Salas RE, Gamaldo CE. Adverse Effects of Sleep Deprivation in the ICU. Crit Care Clin. 2008;24:461–476. Thompson DR, Hamilton DK, Cadenhead CD, et al. Guidelines for intensive care unit design. Crit Care Med. 2012;40:1586–600. DAS-Taskforce 2015, Baron R, Binder A, et al. Evidence and consensus based guideline for the management of delirium, analgesia, and sedation in intensive care medicine. Revision 2015 (DAS-Guideline 2015) - short version. Ger Med Sci. 2015;13:Doc19. Barr J, Fraser GL, Puntillo K, et al. Clinical Practice Guidelines for the Management of Pain, Agitation, and Delirium in Adult Patients in the Intensive Care Unit. Crit Care Med. 2013;41:263–306. National Collaborting Centre for Acute and Chronic Conditions. Delirium: Diagnosis, Prevention and Management. Delirium Diagnosis, Prev Manag. 2010;1–33. Weinhouse GL, Watson PL. Sedation and sleep disturbances in the ICU. Anesthesiol Clin. 2011;29:675–685. Van Rompaey B, Elseviers MM, Van Drom W, et

47.

48.

49.

50.

51.

52.

53.

54.

55.

56.

57.

58.

59.

al. The effect of earplugs during the night on the onset of delirium and sleep perception: a randomized controlled trial in intensive care patients. Crit Care. 2012;16:R73. Gabor JY, Cooper AB, Crombach SA, et al. Contribution of the intensive care unit environment to sleep disruption in mechanically ventilated patients and healthy subjects. Am J Respir Crit Care Med. 2003;167:708–715. Walder B, Francioli D, Meyer J-J, et al. Effects of guidelines implementation in a surgical intensive care unit to control nighttime light and noise levels. Crit Care Med. 2000;28:2242. Tembo AC, Parker V. Factors that impact on sleep in intensive care patients. Intensive Crit Care Nurs. 2009;25:314–322. Bosma KJ, Ranieri VM. Filtering out the noise: Evaluating the impact of noise and sound reduction strategies on sleep quality for ICU patients. Crit Care. 2009;13:151. Friese RS. Sleep and recovery from critical illness and injury: a review of theory, current practice, and future directions. Crit Care Med. 2008;36:697–705. Figueroa-Ramos MI, Arroyo-Novoa CM, Lee KA, et al. Sleep and delirium in ICU patients: A review of mechanisms and manifestations. Intensive Care Med. 2009;35:781–795. Mundigler G, Delle-Karth G, Koreny M, et al. Impaired circadian rhythm of melatonin secretion in sedated critically ill patients with severe sepsis. Crit Care Med. 2002;30:536–540. Olofsson K, Alling C, Lundberg D, Malmros C. Abolished circadian rhythm of melatonin secretion in sedated and artificially ventilated intensive care patients. Acta Anaesthesiol Scand. 2004;48:679–684. Perras B, Meier M, Dodt C. Light and darkness fail to regulate melatonin release in critically ill humans. Intensive Care Med. 2007;33:1954– 1958. Frisk U, Olsson J, En N, Hahn RG. Low melatonin excretion during mechanical ventilation in the intensive care unit. Clin Sci. 2004;107:47–53. Shilo L, Dagan Y, Smorjik Y, et al. Patients in the intensive care unit suffer from severe lack of sleep associated with loss of normal melatonin secretion pattern. Am J Med Sci. 1999;317:278– 81. Paul T, Lemmer B. Disturbance of circadian rhythms in analgosedated intensive care unit patients with and without craniocerebral injury. Chronobiol Int. 2007;24:45–61. Elliott R, McKinley S, Cistulli P, Fien M.

Chapter 1

60.

61.

62.

63.

64.

65.

66.

67.

68. 69.

Characterisation of sleep in intensive care using 24-hour polysomnography: An observational study. Crit Care. 2013;17. Drouot X, Bridoux A, Thille AW, et al. Sleep continuity: A new metric to quantify disrupted hypnograms in non-sedated intensive care unit patients. Crit Care. 2014;18:628. Mills GH, Bourne RS. Do earplugs stop noise from driving critical care patients into delirium? Crit Care. 2012;16:139. MacKenzie DJ, Galbrun LGU. Noise levels and noise sources in acute care hospital wards. Build Serv Eng Res Technol. 2007;28:117–131. Richardson A, Allsop M, Coghill E, Turnock C. Earplugs and eye masks: do they improve critical care patients’ sleep? Nurs Crit Care. 2007;12:278–286. Xie H, Kang J, Mills GH. Clinical review: The impact of noise on patients’ sleep and the effectiveness of noise reduction strategies in intensive care units. Crit Care. 2009;13:208. Simons KS, Laheij RJF, van den Boogaard M, et al. Dynamic light application therapy to reduce the incidence and duration of delirium in intensive-care patients: a randomised controlled trial. Lancet Respir Med. 2016;4:194– 202. Mistraletti G, Umbrello M, Sabbatini G, et al. Melatonin reduces the need for sedation in ICU patients: A randomized controlled trial. Minerva Anestesiol. 2015;81:1298–1310. Castro RA, Angus DC, Hong SY, et al. Light and the outcome of the critically ill: an observational cohort study. Crit Care. 2012;16:R132. Castro R, Angus DC, Rosengart MR. The effect of light on critical illness. Crit Care. 2011;15:218. van den Boogaard M, Schoonhoven L, van der Hoeven JG, et al. Incidence and short-term consequences of delirium in critically ill patients: A prospective observational cohort study. Int J Nurs Stud. 2012;49:775–783.

General introduction

I. The measurement of sleep in the ICU

Chapter 2 The clinical measurement and analysis of sleep

Chapter 2

Introduction Among caregivers in over 500 European and Canadian ICUs, the most commonly used way to assess whether ICU patients are sleeping turned out to be subjective [1]. When representatives of ICUs were asked how to determine whether their patients were asleep, 78% answered that patients should be lying quietly with eyes closed, 66% focused on decreases in blood pressure, and 60% expected slow and regular respiration. Very few ICUs systematically used validated questionnaires or more objective methods to determine when their patients are sleeping. Subjective methods commonly evaluate sleep times and the perceived quality of sleep using questionnaires of varying length and complexity, to be filled in by the study subject or an observer. In contrast, objective methods require either movement to estimate activity levels (actigraphy), or measure physiological modalities that normally change with the transitions between wakefulness and sleep or between individual sleep stages. These modalities include core body temperature, photoplethysmography, electrocardiography (ECG), electroencephalography (EEG), electrooculography (EOG), electromyography (EMG) or secondarily derived features. Although previous studies of sleep in the ICU have provided new insights into disturbed biorhythm and sleep in the ICU, their scope, statistical significance and reliability have thus far been constrained by the logistical challenges of measuring and diagnosing sleep [2–16]. These and other practical complications are often cited as major factors holding back further progress in monitoring, understanding, managing, and preventing sleep disruption in the ICU. Large amounts of ICU acquired data are likely needed to design and test potential interventions to improve our patients’ sleep, and therefore the availability of accurate methods to measure outcomes is crucial. Furthermore, according to several international ICU guidelines continuous real-time information on sleep could be beneficial for optimization of clinical workflows and evidence-based reduction of sleep disruption by environmental factors. However, there is currently no validated or accepted method to monitor or study sleep objectively in critically ill patients. We will therefore review the available tools and techniques for monitoring and studying sleep in the ICU specifically, and their respective advantages and disadvantages in theoretical and practical contexts. For this chapter, original articles and review articles written in English were used. Finally, we will give an overview of promising efforts to better define important characteristics of sleep in groups of critically ill patients.

The clinical measurement and analysis of sleep

Polysomnography (PSG) Ever since Loomis [17] first defined discrete stages of sleep found by EEG recording and Aserinsky and Kleitman [18] first showed the cyclic alternation of REM and NREM sleep, EEG has been the primary tool for objective sleep analysis. The EEG is measured by connecting multiple electrodes to the scalp, and measuring voltage changes between these electrodes over time. These voltage changes are largely caused by synchronized activity of large numbers of neurons in the cerebral cortex, and increase with the number of harmonically firing neurons. When combined with electromyography (EMG), and electrooculography (EOG), the method is called polysomnography. For diagnostic purposes, the setup is often augmented with the measurement of ECG, nasal airflow, leg movements, oxygen saturation, and respiratory excursions. For the analysis of clinically obtained PSG data, the heavily protocoled and time-domain based method of analysis first proposed by Rechtschaffen and Kales (hereafter referred to as the R&K rules) remains by far the most popular [19]. This method has since been optimized to better suit clinical practice and mirror the present state of understanding of sleep, culminating in the Manual for the Scoring of Sleep developed by the American Academy of Sleep Medicine (AASM) [20].

Manual visual scoring of PSG data The visual and manual scoring of sleep relies heavily on useful descriptors of sleep, such as delta wave background activity and specific transients; sleep spindles, saw-tooth waves and K-complexes. Additionally, the level of arousal can also be determined by adding channels that provide muscle (i.e. EMG), ocular (i.e. EOG), and cardiac (i.e. ECG) activity. Although this method was originally designed based on the sleep architecture of young adults, it has since been applied to a variety of types of patients and clinical scenarios. Scoring ICU sleep recordings is not straightforward, however. Most PSG-based studies have reported EEG activity that was not compliant with any single sleep stage as defined by the AASM [3, 9, 21–24]. A recent study in a neurologic ICU reported that 65% of recorded EEG activity could not be scored according to standardized criteria [23]. Other studies reported 23-60% of all measured activity to be unclassifiable, potentially due to septic encephalopathy [9], the effects of benzodiazepines, or a comatose state [3]. The EEG may exhibit generalized slowing, burst suppression, an isoelectric state, or polymorphic delta activity (focal or generalized) [25]. Normally, sleep and wake activity are mutually exclusive and relatively easy to visually distinguish from one another. In ICU recordings of sleep however, intrusions of sleep-like activity in otherwise normal waking EEG are often seen. In the context of the treatment of critically ill patients, the contradictory activity is thought to result from certain types of medication, sedation, or an attempt to relieve sleep pressure after extended wakefulness due to an inability to sleep during previous nights [3, 22, 25]. In part due to the limited applicability of traditional sleep stage definitions, the sparsely reported inter-rater reliability between human scorers is lower for ICU sleep recordings

Chapter 2

than for other patients. Elliott et al. [2] reported a Cohen’s kappa of κ = 0.58-0.68, or ‘reasonable’ to ‘good’ agreement [26], for sleep-wake scoring by two combinations of three manual scorers. Agreement on detailed sleep staging was much lower, with only slight agreement for stage N1 (κ = 0.08-0.12), moderate agreement for stage N2 and REM sleep (κ = 0.55-0.58 and κ = 0.41-0.44, respectively), and slight to good agreement for stage N3/N4 (κ = 0.20-0.76), depending on the combination of manual scorers. Ambrogio et al. compared the agreement between two manual scorers for PSG recordings of ICU patients and control patients [27]. Inter-rater reliability was good (κ = 0.74) for recordings of ambulatory patients, but there was only slight agreement on the scoring of recordings of ICU patients (κ = 0.19).

Table 1. Suggested requirements for scalable and reliable sleep measurement tools for clinical use and research purposes in the ICU Clinical ICU sleep monitoring

ICU sleep trials

Interval

Continuous (30-60s interval, 24/7)

Continuous or regular intervals

Analysis

Real-time (on-line and bedside)

Retrospective analysis (off-line)

Required skill level

Trained ICU nurse

Trained researcher or expert

Outcome

Single number/graph/feature

Full hypnogram and detailed report

Reliability

High intra-patient reliability

High inter- and intra-patient reliability

Scoring resolution

Sleep/wake analysis

Preferably 5-stage scoring

The clinical measurement and analysis of sleep

Automated analysis of PSG data Automatic EEG analysis algorithms designed and tuned specifically for evaluation of sleep are close to being as accurate, or as accurate as, manual scorers [9, 28-30]. Consequently, manual analysis of available PSG data is increasingly augmented with, or replaced by, varying degrees of automated analysis. The adoption of automated analysis for ICU research purposes is low, however, although the exact reasons are unclear. Several currently available algorithms that achieve high agreement by simulating the process used by human scorers are discussed below. Some algorithms rely on a single channel of EEG to classify sleep automatically. Specifically developed to enable large scale screening and investigation of the impact of noise on sleep, the ASEEGA software achieved high agreement when compared to manual consensus scores, with Cohen’s kappa values of 0.72 for 5-stage classification [28]. It extracts both temporal and spectral features from a single channel of EEG and notably tailors the definitions of relevant frequency bands to the recorded activity. This should theoretically increase the applicability of the algorithm to abnormal recordings, such as those obtained from sedated or septic ICU patients, although the accuracy for ICU recordings was not reported. Anderer et al. developed the Somnolyzer 24x7 algorithm to mimic the consensus scores of multiple human experts following the AASM analysis workflow on a large database of EEG recordings of healthy subjects [29]. A detailed performance analysis of this algorithm by Punjabi et al. found a high degree of agreement between manual and automated scoring, which also resulted in a similarly high agreement for secondary sleep quality metrics such as the arousal index, TST and sleep efficiency [30].

Spectral analysis of EEG data With the advent and widespread use of computers in research, new means were developed and used to quantify EEG activity. Particularly the shift to analysis in the frequency domain has produced information about the composition and quality of the sleep beyond the information that time-domain manual sleep stage scoring was able to provide. The EEG can be decomposed into an infinite number of pure sinusoidal components, approximated by the Fast Fourier Transform (FFT). Instead of looking at the sum of all these components over time in epochs of fixed lengths, analogous to for instance the ECG, the observer can now derive the relative contribution of individual frequencies to the recorded EEG. This relative contribution is known as the power spectral density (PSD). Starting at the low frequency, long wavelength end of the spectrum, the delta frequency band is commonly defined to span the 0-4 Hz range. The 4-7 Hz band is generally defined as theta activity, 7-12 Hz to define alpha waves, and 1230 Hz is known as beta waves. Even higher frequency activity is generally combined in the gamma band, encompassing all activity with a frequency upwards of 30 Hz. The most extensively investigated frequency band for sleep is the delta band, encompassing the slow wave activity of the brain which is thought to best reflect the homeostatic process of sleep [31, 32]. Delta waves are enhanced during NREM sleep after

Chapter 2

sleep deprivation [33], and are therefore potentially valuable as a correlate for sleep pressure or sleep intensity. The transformation of these slow waves to the PSD is known as delta power or slow wave activity (SWA). As the need for sleep increases over time, delta power in subsequent sleep bouts increases, making delta power a useful parameter to follow changes in sleep need over time [34]. Consequently, many automated or semi-automated NREM sleep classifiers partially rely on the automated quantification of the absolute or relative delta power [35, 36]. The interpretation of the results of delta activity-based classifiers is not straightforward however. Several commonly used drugs for instance disproportionally influence the delta frequency band of the EEG [32, 37], and the normal distribution of EEG activity across the frequency spectrum depends heavily on age, and even on sex [38, 39]. More fundamentally, subject movement, sweating, and other physiological and physical factors may inadvertently cause low frequency artefacts to unexpectedly occur over time and across patients. Many modern softwares used for clinical analysis of EEG data already include algorithms to provide the human scorer with added information of the EEG activity in specific frequency bands. These softwares often feature semi-automated detection of sleep spindles, slow-waves, eye movements and other relevant transient signals as a timesaving aide during visual analysis.

Machine learning and deep learning Instead of simulating the logic involved in human visual analysis of EEG data, machines are also capable of finding patterns by deriving their own logic from previously acquired and annotated data [40]. This process of machine learning results in a heuristic, predominantly black-box, system that takes a certain input and redirect it through several layers of interconnected artificial neurons that form an artificial neural network (ANN). Neurons and connections that contribute to the correct classification of the training data set are rewarded by a learning algorithm, while neurons and connections that do not are suppressed. Alternatively, other machine learning systems such as support vector machines (SVM) are shaped based on experimental and physiological theory [41]. Overwhelmingly though, newly developed machine learning systems to process EEG for sleep analysis rely on deep learning, a more complex form of neural networks [40, 42–47]. Since PSG data consist of many physiological signals with relatively high levels of noise, data are filtered before being used to train a classifier. Further efforts can be made to reduce the high dimensional dataset to a small set of relevant features or components [48]. Most neural network systems rely on a manual or automated pre-processing stage that identifies relevant features based on known definitions such as those in the AASM manual or more abstract indicators of activity, mobility and complexity of the signals [49]. Given sufficient layers and degrees of freedom these neural networks are however capable of deriving relevant features without supervision or annotation, which could greatly benefit large scale application in the ICU [50]. These deep learning networks are

The clinical measurement and analysis of sleep

heralded as the basis for future end-to-end artificially intelligent systems, and have so far demonstrated a remarkable ability to classify a wide range of EEG signals and disorders without human intervention [40]. Deep learning systems trained on EEG data acquired in adult critically ill patients have not been reported on yet despite the promising ability to define features and definitions dynamically depending on the characteristics of the dataset.

Depth of general anaesthesia monitoring systems Besides similarities in behavioural characteristics, some EEG changes that occur as an effect of administering anaesthetics also resemble natural sleep [51]. Particularly the cyclic transition from high frequency low amplitude EEG patterns during wakefulness, to more synchronized and low frequency patterns during deep sleep resembles the transition seen during gradually deepening general anaesthesia. Practical and automated methods currently in use to process EEG to assess depth of general anaesthesia in the operating room (OR), during transport, or in the ICU may therefore be of some practical value in estimating depth of sleep. We will discuss the most common of these systems and the efficacy of their complex parameters derived from several features in the time and frequency domain for use in sleep analysis. One such multivariate parameter is the automatically calculated Bispectral Index (BIS). Sleigh et al. [52] were perhaps the first to describe the BIS as a marker for sleep depth in a small sample. They note that changes in the depth of natural sleep are reflected by changes in the BIS. Gimenez [53] concluded that BIS could provide a useful measure of sleep depth in intensive care units, despite considerable overlap for BIS values with multiple sleep stages. They were also unable to discriminate between REM and NREM stages using the BIS. Interestingly, BIS values decreased after 40h of sleep deprivation. Similarly, Vacas et al. [54] evaluated the merits of the Sedline brain monitor compared to manual analysis in a sleep lab and ICU patients. They were able to achieve good agreement for wake and NREM stage 2, but found low agreement for NREM stage 1 and 3. Nieuwenhuijs et al. [55] compared the BIS and the spectral edge frequency (SEF) to manual analysis and found similar overlap for BIS and 95% SEF values with different natural sleep stages. Dahaba et al. [56] and Bennisa et al. [57] tested a newer version of the BIS monitor on sleep deprived anaesthesiologists and young patients, respectively. Both found similar, although reduced, overlap for BIS values with multiple stages of sleep. They concluded that the BIS showed a temporal decline that corresponded with progressively deepening sleep. Due to the lack of information about muscle activity, they were unable to discriminate between wakefulness, N1 and REM sleep. Tung et al. [58] in contrast, found that the BIS could be used to detect the onset of natural sleep, and specifically suggested the use of the BIS as a monitor for inadvertent sleep onset. Generally, currently available processed EEG devices are unable to make a distinction between natural sleep and chemically induced sleep or coma, by design. For accurate scoring of natural sleep, specific information from the spectral and time domain is needed. Although the general slowing of EEG patterns is similar for general anaesthesia

Chapter 2

and natural sleep, natural sleep can be further divided into cyclically alternating stages with unique time domain transient waveforms, spatiotemporal characteristics, and changes in muscle tone and eye movements. To correct for these differences between BIS and traditional sleep staging Nicholson et al. [59] augmented BIS with two EMG leads and defined manual criteria for three different macro sleep structures: relatively normal cyclical sleep, abnormal sleep, and an overall lack of sleep. Using this division, they were able to make some statements on the distribution of severely disturbed EEG patterns among ICU patients, but did not report on the agreement with other sleep scoring methods. Regardless of their accuracy, processed EEG indices are dimensionless and therefore hard to interpret without a clear understanding of their origins. This further complicates their use in a population where hypnotics and sedative medication are commonly used. These limitations of processed EEG devices do however not detract from their practical and logistical advantages, such as easy and quick access to EEG data, practical form-factors with minimal setup time, and a limited number of output indices that are comparable between studies.

Actigraphy Actigraphs are small, often wrist worn devices that use accelerometers to determine movement along multiple axes. When multiple days of recordings are available, the average gross motor activity of a subject can be derived in a very non-invasive and unobtrusive fashion. Shilo et al. first used actigraphy in 1999 to compare ICU patients’ sleep distribution to that of general medical ward control patients. They reported that ICU patients only sleep for short periods of time based on the actigraphy, but did not comment on the usability or reliability of the used actigraph [60]. Beecroft et al. studied 12 ICU patients and compared actigraphy to polysomnography using fixed thresholds for wake and sleep in the actigraph and found an overall agreement of up to 61% [61]. The correlation between secondary sleep quantity and quality measures such as total sleep time and sleep efficiency was low, however. They concluded that actigraphy is an inaccurate and unreliable alternative to PSG in the ICU population, likely due to the inability to distinguish between voluntary movements and nursing care activities, and between sleep and motionless wakefulness.

Alternative physiological parameters Sleep influences many physiological parameters under control of the autonomic nervous system, such as heart rate and blood pressure. An often used marker for autonomic function is heart rate variability (HRV), the beat-to-beat difference in heart rate, which can be derived from ECG alone. HRV is long known to change with transitions from the sympathetically dominated REM sleep, to the parasympathetically dominated NREM sleep stages [62]. For this reason, it has repeatedly been investigated as an easy to administer beat-to-beat estimate of sleep stage, although it was only used as a basic parameter for autonomic function in non-surgical patients [63–65]. Similarly, the pulse

The clinical measurement and analysis of sleep

pressure variation (PPV) or core body temperature could potentially provide a surrogate for general autonomic function, although these too were not tested as a parameter for the estimation of the depth of sleep in the ICU population. Furthermore, the degree of interaction between autonomic nervous cardiac and respiratory activity, known as high-frequency cardiopulmonary coupling (CPC), has been suggested as a biomarker for stable sleep [66]. Thomas et al. found a strong correlation between EEG delta power and CPC in a large database of home-based PSG recordings, independent of NREM sleep stages. This tight integration of neural and cardiopulmonary control could provide an opportunity to find estimates of deep sleep or sedation from ECG derived parameters, but has not been investigated as such.

Questionnaires and sleep scales Subjective sleep scales and questionnaires are perhaps the most controversial and widely used tools in ICU sleep research, particularly for interventional studies. Due to the practical constraints of working with ICU patients, these tools were designed specifically for quick and repeated assessment of perceived sleep on the ICU. The most popular examples developed specifically for use in the ICU are the five-item Richards-Campbell Sleep Questionnaire (RCSQ) [67], and the 27 item Sleep in the Intensive Care Unit Questionnaire (SICQ) [68]. By far the most commonly used sleep questionnaire to evaluate perceived sleep in the critically ill is the RCSQ [2, 69–77]. Developed and validated by Richards et al., it is comprised of five visual analogue scales, one for each of five domains: overall sleep depth, sleep latency, number of awakenings, percentage of time awake and the overall quality of sleep [67]. An overall sleep index is calculated from the mean score of these five questions. Sometimes adapted to suit specific requirements [78], it is generally deemed a quick and easy way for patients and nurses to determine perceived sleep quality of ICU patients. The RCSQ is generally reported to have high internal consistency, with a Cronbach’s alpha coefficient ranging from 0.82 to 0.92 [67, 69, 71] for selfassessment, and 0.83 to 0.95 [67, 69] for nurse assessment. One obvious limitation of the use of questionnaires in an environment where patients are often anxious, confused, exhausted, sedated or even comatose, is the limited applicability for self-assessment. Frisk et al. found that out of a sample of 31 surgical ICU patients that stayed at least two nights, approximately half (15, 48%) were able to fill out the RCSQ [69]. Bourne et al. found that 80% of a group of 24 patient with acute respiratory failure were able to complete the RCSQ [21]. Despite RCSQ scores varying widely between patients, both Richards et al. and Frisk et al. found that there was no significant difference between patient and nurse assessment of perceived ICU sleep [67, 69]. Conversely, Nicolás et al. found only 50% agreement when asking surgical ICU patients to rate their sleep on the RCSQ by verbally rating each element, and found that nurses generally overestimated patients’ sleep in the other half of the observations [75]. Kamdar et al. provided a test of patient-nurse inter-rater reliability

Chapter 2

using the RCSQ over 92 patient days, and similarly concluded that there was only slight to moderate agreement with nurses overestimating their patients’ sleep in the seven categories of the instrument (intra-class correlation coefficient = 0.14-0.49) [77]. Regardless of the agreement with other more objective or continuous methods of quality of sleep estimation, the subjective experience of the patient seems relevant. Where more objective methods are often far removed from immediate clinical relevance, although perhaps better suited for investigative purposes, the subjective experience of the patient directly reflects the degree to which sleep was perceived to be disturbed. Furthermore, the ability to easily and cheaply determine day-to-day changes in perceived quality of sleep widens the scope and duration of any study of sleep. It may only be applicable to conscious, clear-minded and communicative patients, but these may also be the patients most likely to benefit from practical interventions to improve quality of sleep. Furthermore, sleep scales and questionnaires currently also represent the only feasible way to perform large scale trials, and may be administered without expert knowledge or experience.

The clinical measurement and analysis of sleep

Quality and quantity of sleep Quantity of sleep Determining the quantity of sleep after scoring is fairly straightforward for individual patients, although very little is known about the optimal quantity of sleep during critical illness. It is assumed that the long periods of normal nocturnal sleep may be preferable over short bouts of fragmented sleep evenly distributed over day and night seen during critical illness, although there is little supporting scientific evidence. Sleep is intuitively associated with health, and efficient sleep is known to improve defences against a wide range of infections [79], although it is largely unknown how exactly sleep helps recovery from illness, and how much sleep is required. In absence of sickness, the optimal quantity of sleep largely depends on age, but may vary significantly between individuals of the same age [31, 80, 81]. The atypical EEG patterns that replace the usual sleep and wake EEG activity in critically ill patients have encouraged the development of new pragmatic scoring rules. Drouot et al. defined two new EEG classes to allow for atypical patterns found in the ICU to be scored: ‘atypical sleep’ and ‘pathologic wakefulness’ [22]. Watson et al. then defined a new decision tree for ICU sleep scoring, with the addition of the new ‘atypical sleep’ class [82]. Although it is debatable whether the observed atypical EEG activity constitutes sleep, the dissociation from traditional sleep stages seems warranted and valuable [83]. The total sleep time (TST) is normally defined as the unweighted sum of all sleep activity during a 24-hour day, or during one night. In other studies, it is only defined over a specific timeframe where the subject intends to sleep, from lights off till lights on, known as total bed time (TBT). Over the same time period, sleep efficiency can be calculated as the percentage of intended sleep time actually spent asleep. The total sleep period (TSP) is defined as the time between the first sleep activity and the last awakening during the recording. All wake activity in between is called wake after sleep onset (WASO). Sleep latency is the delay between light off and the first sleep activity, REM latency is similarly defined as the time between the first sleep activity and the first REM sleep activity. Some of these metrics (TBT, sleep efficiency, sleep latency) have no clear meaning in ICU settings, where most patients are confined to their bed and a formal lights on or off time is hard to define. These metrics are commonly calculated with a pragmatic lights on/off time in ICU research. A recent study investigated the efficacy of a new metric for sleep continuity, classifying segments of continuous sleep as either ‘bouts’, ‘short naps’, or ‘long naps’ [83]. This approach is largely based on the concept of a lower threshold for recuperative sleep. Continuous sleep exceeding 10 minutes in duration was previously shown to be more recuperative than shorter bouts [84]. This metric may provide a more detailed estimation of the amount of potentially recuperative sleep in ICU patients than existing metrics such as the arousal index and TST. The presence of a lower threshold for continuity required for recuperative sleep is perhaps explained by the relatively quick formation of new connections between dendrites, as demonstrated in mice after motor learning tasks [85].

Chapter 2

Regardless of whether or not plasticity is a limiting factor in achieving recuperative sleep, there is a pressing need for robust, automated and detailed methods to determine quality of sleep despite disturbed sleep patterns.

Quality of sleep The quality of sleep is a commonly used, but very loosely defined term to describe the ability of a period of sleep to replenish cognitive capacity and optimize efficiency. The definition of quality of sleep should include quantitative aspects of a period of sleep known to influence objective and subjective outcomes, such as TST and continuity [84, 86]. Qualitative aspects, often determined subjectively through questionnaires and cognitive performance metrics, may also be included in the evaluation of quality of sleep. These include general indices for how well rested and recuperated patients feel after a period of sleep. In general, sleep quality seems to be best defined by subjective experience and directly measuring relative performance after a period of sleep. Alternatively, it can be determined indirectly by detailed analysis of the EEG during interrupted nights. The observed EEG activity can then be correlated to known effects on clinically relevant performance outcomes in previous samples, such as incidence of delirium, delirium free days, length of stay, etc. Most ICU studies forgo these clinical measurements of relative clinical performance after sleep, and estimate the quality of sleep by the percentage per stage of sleep, or by the average duration of an episode of sleep instead [2]. Bourne et al. [87] used the area under the curve (AUC) of the BIS trace to provide an indication of both sleep quantity and sleep quality, or more specifically the cumulative sleep depth, thereby assuming deeper stages of sleep to be linearly more clinically relevant. They state that although the clinical significance and interpretation of the AUC of the BIS trace is unclear, it might be more informative than sleep quantity alone. Specific optional PSG modalities such as nasal airflow, leg EMG, exhaled CO2 and the indices derived from these features are rarely reported in ICU populations. The additional electrodes needed to determine these features are often dropped in favour of an easier and less intrusive setup, and the incidence of these features is rarely reported. These pragmatic methods suggest that quality of sleep is best estimated by a weighted sum of the stages of sleep apparent in an episode of sleep. The potential importance of the exact sequence and duration of sleep stages, the so-called sleep architecture, is ignored because the translation to clinically relevant outcomes is unknown. Generally, even in healthy subjects there is much debate on what exactly constitutes normal or good sleep, with even greater uncertainty in ill or critically ill populations. An argument could be made for the expansion of the used methods to obtain as much information on the patients’ physiology and environment during and after sleep as possible, to determine the clinical importance of specific sleep features. However, small scale pilot studies in the ICU have so far been unable to produce or detect clinically significant changes in patient outcomes such as mortality or length of stay, due to the scope limiting and labour-intensive practice of PSG and the inability of subjective measurements to accurately quantify sleep disruption.

The clinical measurement and analysis of sleep

Discussion The overall lack of a validated and practical standard is perhaps the most debilitating factor in the current state of ICU sleep research. This overview examined available alternatives and techniques used in mathematics, computer science, neuroscience and behavioural sciences that have only recently started being experimented with in the field of intensive care. The flawed standard approach of PSG with manual scoring has proven to be cumbersome, expensive, time-consuming and in many cases incompatible with unresolvable EEG activity found in the ICU. Accuracy and agreement between scorers depend heavily on patient specific differences for most available methods used in sleep research. The need for adaptive algorithms seems to be even larger for ICU recordings, where medication and sedation may also play a part in increasing inter-patient differences. Given the resulting large inter- and intra-patient differences, rule-based systems could in the future be augmented or replaced by neural networks for deeplearning capabilities, or more simple interpretable spectral indices for longitudinal studies. So far, no fully automated AASM analysis of ICU recorded PSG data has been reported on. In general, the available algorithms are very capable of extracting the required features from normal EEG recordings and applying the various logical rules set by the AASM. The resulting agreement with any individual human scorer not included in the training set is, however, often lower due to differences between the exact methods being employed by the human scorers. The unintended smoothing and dynamic thresholding that is applied to varying degrees by human scorers makes it impossible to increase the accuracy of trained algorithms beyond the inter-rater agreement found between human scorers. For the accuracy of an algorithm to exceed that of a pooled training set, it would need to be able to mimic scorers individually, by changing thresholds and the level of smoothing as required within the limits set by the AASM. Inter-rater reliability is often expressed as Cohen’s Kappa statistic, although the interpretation is not straight-forward. Large classimbalances exist in recordings of normal sleep, which are exacerbated in ICU recordings, where REM and SWS are rarely seen. This may further reduce the inter-rater agreement expressed as the Kappa statistic, but also the significance of the Kappa statistic itself as a measure for agreement. The EEG of ICU patients is confounded by not well understood encephalopathic patterns that may further hamper the analysis of sleep. Vice versa, EEG may be the only modality to provide insights in the multifactorial aetiology of these observed patterns. Other techniques such as questionnaires are perhaps more suited to validate the effects of new interventions to improve sleep, but they may be less useful in determining these underlying causes for disturbed sleep in the individual patient. Obtaining high quality EEG data is further complicated by frequent and intensive care interventions, profuse sweating or high motility of patients. This is where the easy to (re)apply self-adhering electrodes of the BIS or other processed EEG systems may

Chapter 2

provide an example to reduce the fail-rate and increase the longevity of ICU sleep monitoring. Recent developments in active [88] and dry [89] electrode designs may provide future analysis systems with the needed practical implementation, but have not yet been proven in the demanding ICU environment. A potential shortcoming of methods relying on continuous measurement of the EEG is that the attached electrodes may impact the quality of sleep. Furthermore, body posture [90] may modify the efficiency of sleep but is hampered by the restrictive setup needed for PSG. Further investigation of more comfortable or minimal ways of EEG recording seems warranted. Most studies are performed on surgical ICUs, where a relatively homogenous group of patients is admitted for relatively short periods of time. Perhaps patients with long admission times have the most to gain from structural interventions, since they are most likely to develop significant problems due to long exposure to the ICU environment. In these cases, a simple spectral EEG index or binary analysis system may suffice to screen for sleep deprivation and optimally schedule interventions around patients’ attempts to sleep. Depending on the goals of a trial, the RCSQ and similar validated questionnaires may provide relevant insights in perceived sleep quality when patients are capable of evaluating it themselves. Caution has to be taken to not overestimate sleep when it is evaluated by nurses instead of patients themselves. Furthermore, the correlation between perceived sleep quality and objectively measured sleep in ICU populations is weak, so simultaneous objective measurement seems recommendable in most cases.

The clinical measurement and analysis of sleep

Future perspectives The development of a practical method of measuring, analysing, understanding and eventually improving sleep in the ICU is hampered by the lack of large datasets, convincing clinical importance, and setups that are easy to apply and maintain. A combination of rapid developments in dynamic and unsupervised learning algorithms, innovations in electrode design, and ever-growing computational power drive by the consumer market seems best suited for large scale application in the ICU. Although fundamental research on the topic likely requires expensive multi-channel recordings and 5-stage sleep scoring, sleep monitoring for scheduling of daily care around periods of sleep may only require a single channel sleep/wake system. Although the measurement of sleep in the ICU is currently largely clinically inconsequential due to scale limitations, it may in the future lead to the development of tailored interventions to help facilitate sleep for those that most need it. Furthermore, a clearer understanding of the wide range of potentially sleep disrupting factors may determine whether interventions are deemed beneficial on a per patient basis. Not all patients can be expected to benefit from environmental optimizations, but those who can are likely found in the ICU.

Chapter 2

References 1.

10.

11.

12. 13.

14.

15.

16.

Hofhuis JGM, Rose L, Blackwood B, et al. Clinical practices to promote sleep in the ICU: A multinational survey. Int J Nurs Stud. 2018;81:107–114. Elliott R, McKinley S, Cistulli P, Fien M. Characterisation of sleep in intensive care using 24-hour polysomnography: an observational study. Crit Care. 2013;17:R46. Cooper AB, Thornley KS, Young GB, et al. Sleep in critically ill patients requiring mechanical ventilation. Chest. 2000;117:809–818. Friese RS. Sleep and recovery from critical illness and injury: a review of theory, current practice, and future directions. Crit Care Med. 2008;36:697–705. Beltrami FG, Nguyen XL, Pichereau C, et al. Sono na unidade de terapia intensiva. J Bras Pneumol. 2015;41:539–546. Tembo AC, Parker V. Factors that impact on sleep in intensive care patients. Intensive Crit Care Nurs. 2009;25:314–322. Andersen JH, Boesen HC, Skovgaard Olsen K. Sleep in the Intensive Care Unit measured by polysomnography. Minerva Anestesiol. 2013;79:804–15. Kamdar BB, Needham DM, Collop NA. Sleep Deprivation in Critical Illness: Its Role in Physical and Psychological Recovery. J Intensive Care Med. 2012;27:97–111. Freedman NS, Gazendam J, Levan L, et al. Abnormal sleep/wake cycles and the effect of environmental noise on sleep disruption in the intensive care unit. Am J Respir Crit Care Med. 2001;163:451–7. Drouot X, Cabello B, d’Ortho M-P, Brochard L. Sleep in the intensive care unit. Sleep Med Rev. 2008;12:391–403. Korompeli A, Muurlink O, Kavrochorianou N, et al. Circadian disruption of ICU patients: A review of pathways, expression, and interventions. J Crit Care. 2017;38:269–277. Wang J, Greenberg H. Sleep and the ICU. Open Crit Care Med J. 2013;6:80–87. Pulak LM, Jensen L. Sleep in the Intensive Care Unit: A Review. J Intensive Care Med. 2016;31:14–23. Figueroa-Ramos MI, Arroyo-Novoa CM, Lee KA, et al. Sleep and delirium in ICU patients: A review of mechanisms and manifestations. Intensive Care Med. 2009;35:781–795. Weinhouse GL, Schwab RJ, Watson PL, et al. Bench-to-bedside review: Delirium in ICU patients - importance of sleep deprivation. Crit Care. 2009;13:234. Litton E, Carnegie V, Elliott R, Webb SAR. The

17.

18.

19.

20.

21.

22.

23.

24.

25.

26.

27.

28.

29.

30.

Efficacy of Earplugs as a Sleep Hygiene Strategy for Reducing Delirium in the ICU. Crit Care Med. 2016;Publish Ah:1. Loomis AL, Harvey EN, Hobart GA. Cerebral states during sleep, as studied by human brain potentials. J Exp Psychol. 1937;21:127–144. Aserinsky E, Kleitman N. Regularly occurring periods of eye motility, and concomitant phenomena, during sleep. Science. 1953;118:273–4. Kales A, Rechtschaffen A. A Manual of Standardized Terminology, Techniques and Scoring System for Sleep Stages of Human Subjects. U.S. National Institute of Neurological Diseases and Blindness, Neurological Information Network, Bethesda, Md. Iber CC, Ancoli-israel S, Chesson AL, et al. The AASM manual for the scoring of sleep and associated events; rules, terminology and technical specifications, 1st ed. American Academy of Sleep Medicine, Westchester, IL. Bourne RS, Minelli C, Mills GH, Kandler R. Clinical review: Sleep measurement in critical care patients: research and clinical implications. Crit Care. 2007;11:226. Drouot X, Roche-Campo F, Thille AW, et al. A new classification for sleep analysis in critically ill patients. Sleep Med. 2012;13:7–14. Foreman B, Westwood AJ, Claassen J, Bazil CW. Sleep in the Neurological Intensive Care Unit. J Clin Neurophysiol. 2015;32:66–74. Watson PL. Measuring sleep in critically ill patients: beware the pitfalls. Crit Care. 2007;11:159. Weinhouse GL, Watson PL. Sedation and sleep disturbances in the ICU. Anesthesiol Clin. 2011;29:675–685. Landis JR, Koch GG. The measurement of observer agreement for categorical data. Biometrics. 1977;33:159–174. Ambrogio C, Koebnick J, Quan SF, et al. Assessment of sleep in ventilator-supported critically III patients. Sleep. 2008;31:1559–1568. Berthomier C, Drouot X, Herman-Stoïca M, et al. Automatic analysis of single-channel sleep EEG: validation in healthy individuals. Sleep. 2007;30:1587–1595. Anderer P, Moreau A, Woertz M, et al. Computer-Assisted Sleep Classification according to the Standard of the American Academy of Sleep Medicine: Validation Study of the AASM Version of the Somnolyzer 24 × 7. Neuropsychobiology. 2010;62:250–264. Punjabi NM, Shifa N, Dorffner G, et al. Computer-Assisted Automated Scoring of

The clinical measurement and analysis of sleep

31. 32.

33.

34.

35.

36.

37.

38.

39.

40.

41.

42.

43.

44.

Polysomnograms Using the Somnolyzer System. Sleep. 2015;38:1555–1566. Feinberg I. Changes in sleep cycle patterns with age. J Psychiatr Res. 1974;10:283–306. Davis CJ, Clinton JM, Jewett KA, et al. Delta wave power: An independent sleep phenotype or epiphenomenon? J Clin Sleep Med. 2011;7:S16-8. Pappenheimer JR, Koski G, Fencl V, et al. Extraction of sleep-promoting factor S from cerebrospinal fluid and from brains of sleepdeprived animals. J Neurophysiol. 1975;38:1299– 311. Akerstedt T, Gillberg M, Folkard S. Slow wave activity and prior sleep/wakefulness on an irregular schedule. J Sleep Res. 1992;1:118-121. Šušmáková K, Krakovská A. Discrimination ability of individual measures used in sleep stages classification. Artif Intell Med. 2008;44:261–277. Brignol A, Al-ani T, Drouot X. Phase space and power spectral approaches for EEG-based automatic sleep-wake classification in humans: A comparative study using short and standard epoch lengths. Comput Methods Programs Biomed. 2013;109:227–238. Bourne RS, Mills GH. Sleep disruption in critically ill patients - Pharmacological ocnsiderations. Anaesthesia. 2004;59:374–384. Scher MS. Ontogeny of Eeg Sleep from Neonatal through Infancy Periods. Sleep A Compr Handb. 2005;9:487–506. Reynolds CF, Kupfer DJ, Thase ME, et al. Sleep, gender, and depression: An analysis of gender effects on the electroencephalographic sleep of 302 depressed outpatients. Biol Psychiatry. 1990;28:673–684. Movahedi F, Coyle JL, Sejdic E. Deep belief networks for electroencephalography: A review of recent contributions and future outlooks. IEEE J Biomed Heal Informatics. 2017;1–1. Singla R. Comparison of SVM and ANN for classification of eye events in EEG. J Biomed Sci Eng. 2011;04:62–69. Ronzhina M, Janoušek O, Kolářová J, et al. Sleep scoring using artificial neural networks. Sleep Med Rev. 2012;16:251–263. Schwaibold M, Schöller B, Penzel T, Bolz A. Artificial intelligence in sleep analysis (ARTISANA) - Modelling of the visual sleep stage identification process . Artif Intell sleep Anal - Model des Vis Vor bei der schlafklassifikation. 2001;46:129–132. Mamelak AN, Quattrochi JJ, Allan Hobson J. Automated staging of sleep in cats using neural

45. 46.

47.

48.

49.

50.

51. 52.

53.

54.

55.

56.

57.

58.

networks. Electroencephalogr Clin Neurophysiol. 1991;79:52–61. Chan M-C, Spieth PM, Quinn K, et al. Circadian rhythms. Crit Care Med. 2012;40:246–253. Zhang J, Wu Y. A New Method for Automatic Sleep Stage Classification. IEEE Trans Biomed Circuits Syst. 2017;1–14. Supratak A, Dong H, Wu C, Guo Y. DeepSleepNet: a Model for Automatic Sleep Stage Scoring based on Raw Single-Channel EEG. IEEE Trans Neural Syst Rehabil Eng. 2017;4320:1–11. Boostani R, Karimzadeh F, Nami M. A comparative review on sleep stage classification methods in patients and healthy individuals. Comput Methods Programs Biomed. 2017;140:77–91. Najdi S, Gharbali AA, Fonseca JM. Feature ranking and rank aggregation for automatic sleep stage classification: a comparative study. Biomed Eng Online. 2017;16:78. Bengio Y, Lecun Y. Scaling Learning Algorithms towards AI. In: Large Scale Kernel Machines. pp 321–360. Musizza B, Ribaric S. Monitoring the Depth of Anaesthesia. Sensors. 2010;10:10896–10935. Sleigh JW, Andrzejowski J, Steyn-Ross a, Steyn-Ross M. The bispectral index: a measure of depth of sleep? Anesth Analg. 1999;88:659– 661. Giménez S, Romero S, Alonso JF, et al. Monitoring sleep depth: analysis of bispectral index (BIS) based on polysomnographic recordings and sleep deprivation. J Clin Monit Comput. 2015;1–8. Vacas S, McInrue E, Gropper MA, et al. The Feasibility and Utility of Continuous Sleep Monitoring in Critically Ill Patients Using a Portable Electroencephalography Monitor. Anesth Analg. 2016;123:13–15. Nieuwenhuijs D, Coleman EL, Douglas NJ, et al. Bispectral index values and spectral edge frequency at different stages of physiologic sleep. Anesth Analg. 2002;94:125–129. Dahaba AA, Xue JX, Xu GX, et al. Bilateral bispectral index (BIS)-vista as a measure of physiologic sleep in sleep-deprived anesthesiologists. Minerva Anestesiol. 2011;77:388–393. Benissa MR, Khirani S, Hartley S, et al. Utility of the bispectral index for assessing natural physiological sleep stages in children and young adults. J Clin Monit Comput. Tung A, Lynch JP, Roizen MF. Use of the bis monitor to detect onset of naturally occurring

Chapter 2

sleep. J Clin Monit Comput. 2001;17:37–42. 59. Nicholson T, Patel J, Sleigh JW. Sleep Patterns in Intensive Care Unit Patients : A Study Using the Bispectral Index. Crit. care Resusc. 3:86–91. 60. Shilo L, Dagan Y, Smorjik Y, et al. Patients in the intensive care unit suffer from severe lack of sleep associated with loss of normal melatonin secretion pattern. Am J Med Sci. 1999;317:278– 81. 61. Beecroft JM, Ward M, Younes M, et al. Sleep monitoring in the intensive care unit: Comparison of nurse assessment, actigraphy and polysomnography. Intensive Care Med. 2008;34:2076–2083. 62. Snyder F, Hobson JA, Morrison DF, Goldfrank F. Changes in Respiration, Heart Rate, and Systolic Blood Pressure in Human Sleep. J Appl Physiol. 1964;19:417–422. 63. Winchell RJ, Hoyt DB. Spectral Analysis of Heart Rate Variability in the ICU: A Measure of Autonomic Function. J Surg Res. 1996;63:11–16. 64. Pontet J, Contreras P, Curbelo A, et al. Heart rate variability as early marker of multiple organ dysfunction syndrome in septic patients. J Crit Care. 2003;18:156–63. 65. Kasaoka S, Nakahara T, Kawamura Y, et al. Real-time monitoring of heart rate variability in critically ill patients. J Crit Care. 2010;25:313–316. 66. Thomas RJ, Mietus JE, Peng C-K, et al. Relationship between delta power and the electrocardiogram-derived cardiopulmonary spectrogram: possible implications for assessing the effectiveness of sleep. Sleep Med. 2014;15:125–31. 67. Richards KC, O’Sullivan PS, Phillips RL. Measurement of sleep in critically ill patients. J Nurs Meas. 2000;8:131–44. 68. Freedman NS, Kotzer N, Schwab RJ. Patient perception of sleep quality and etiology of sleep disruption in the intensive care unit. Am J Respir Crit Care Med. 1999;159:1155–1162. 69. Frisk U, Nordström G. Patients’ sleep in an intensive care unit—patients’ and nurses’ perception. Intensive Crit Care Nurs. 2003;19:342–349. 70. Li SY, Wang TJ, Vivienne Wu SF, et al. Efficacy of controlling night-time noise and activities to improve patients’ sleep quality in a surgical intensive care unit. J Clin Nurs. 2011;20:396–407. 71. Ritmala-Castren M, Axelin A, Kiljunen K, et al. Sleep in the intensive care unit - nurses’ documentation and patients’ perspectives. Nurs Crit Care. 2014. 72. Krotsetis S, Richards KC, Behncke A, Köpke S. The reliability of the German version of the Richards Campbell Sleep Questionnaire. Nurs

Crit Care. 2017;1–6. 73. Aitken LM, Elliott R, Mitchell M, et al. Sleep assessment by patients and nurses in the intensive care: An exploratory descriptive study. Aust Crit Care. 2016;1–8. 74. McKinley S, Fien M, Elliott R, Elliott D. Sleep and psychological health during early recovery from critical illness: an observational study. J Psychosom Res. 2013;75:539–45. 75. Nicolás A, Aizpitarte E, Iruarrizaga A, et al. [Perception of night-time sleep by the surgical patients in an intensive care unit]. Enferm intensiva. 2008;13:57–67. 76. Williamson JW. The effects of ocean sounds on sleep after coronary artery bypass graft surgery. Am J Crit Care. 1992;1:91–7. 77. Kamdar BB, Shah PA, King LM, et al. Patientnurse inter-rater reliability and agreement of the Richards-Campbell sleep questionnaire. Am J Crit Care. 2012;21:261–9. 78. Stewart JA, Green C, Stewart J, Tiruvoipati R. Factors influencing quality of sleep among nonmechanically ventilated patients in the Intensive Care Unit. Aust Crit Care. 2016;PublishAh:7–9. 79. Imeri L, Opp MR. How (and why) the immune system makes us sleep. Nat Rev Neurosci. 2009;10:199–210. 80. Bonnet MH, Arand DL. EEG arousal norms by age. J Clin Sleep Med. 2007;3:271–4. 81. Lancee J, Eisma MC, van Straten A, Kamphuis JH. Sleep-Related Safety Behaviors and Dysfunctional Beliefs Mediate the Efficacy of Online CBT for Insomnia: A Randomized Controlled Trial. Cogn Behav Ther. 2015;44:406– 422. 82. Watson PL, Pandharipande P, Gehlbach BK, et al. Atypical Sleep in Ventilated Patients. Crit Care Med. 2013;41:1958–1967. 83. Drouot X, Bridoux A, Thille AW, et al. Sleep continuity: A new metric to quantify disrupted hypnograms in non-sedated intensive care unit patients. Crit Care. 2014;18:628. 84. Downey R, Bonnet MH. Performance during frequent sleep disruption. Sleep. 1987;10:354– 363. 85. Yang G, Gan WB. Sleep contributes to dendritic spine formation and elimination in the developing mouse somatosensory cortex. Dev Neurobiol. 2012;72:1391–1398. 86. Williamson a M, Feyer a M. Moderate sleep deprivation produces impairments in cognitive and motor performance equivalent to legally prescribed levels of alcohol intoxication. Occup Environ Med. 2000;57:649–655. 87. Bourne RS, Mills GH, Minelli C. Melatonin

The clinical measurement and analysis of sleep

therapy to improve nocturnal sleep in critically ill patients: encouraging results from a small randomised controlled trial. Crit Care. 2008;12:R52. 88. Xu J, Mitra S, Van Hoof C, et al. Active Electrodes for Wearable EEG Acquisition: Review and Electronics Design Methodology. IEEE Rev Biomed Eng. 2017;10:187–198. 89. Lopez-Gordo MA, Sanchez-Morillo D, Pelayo Valle F. Dry EEG electrodes. Sensors (Basel). 2014;14:12847–70. 90. Lee H, Xie L, Yu M, et al. The Effect of Body Posture on Brain Glymphatic Transport. J Neurosci. 2015;35:11034–11044. .

Chapter 3 Intensive care unit depth of sleep: Proof of concept of a simple electroencephalography index in the non-sedated Critical Care 2014 18(2) R66 Laurens Reinke, Johannes H. van der Hoeven, Michel J.A.M van Putten, Willem Dieperink, Jaap E. Tulleken

Chapter 3

Abstract Introduction Intensive care unit (ICU) patients are known to experience severely disturbed sleep, with possible detrimental effects on short- and long- term outcomes. Investigation into the exact causes and effects of disturbed sleep has been hampered by cumbersome and time-consuming methods of measuring and staging sleep. We introduce a novel method for ICU depth of sleep analysis, the ICU depth of sleep index (IDOS index), using single channel electroencephalography (EEG) and apply it to outpatient recordings. A proof of concept is shown in non-sedated ICU patients.

Methods Polysomnographic (PSG) recordings of five ICU patients and 15 healthy outpatients were analysed using the ICU depth of sleep (IDOS) index, based on the ratio between gamma and delta band power. Manual selection of thresholds was used to classify data as either wake, sleep or slow wave sleep (SWS). This classification was compared to visual sleep scoring by Rechtschaffen & Kales criteria in normal outpatient recordings and ICU recordings to illustrate face validity of the IDOS index.

Results When reduced to two or three classes, the scoring of sleep by IDOS index and manual scoring show high agreement for normal sleep recordings. The obtained overall agreements, as quantified by the kappa coefficient, were 0.84 for sleep/wake classification and 0.82 for classification into three classes (wake, non-SWS and SWS). Sensitivity and specificity were highest for the wake state (93% and 93%, respectively) and lowest for SWS (82% and 76%, respectively). For ICU recordings, agreement was similar to agreement between visual scorers previously reported in literature.

Conclusions Besides the most satisfying visual resemblance with manually scored normal PSG recordings, the established face-validity of the IDOS index as an estimator of depth of sleep was excellent. This technique enables real-time, automated, single channel visualization of depth of sleep, facilitating the monitoring of sleep in the ICU.

Intensive care unit depth of sleep Proof of concept of a simple electroencephalography index in the non-sedated

Introduction Sleep is a dynamic, complex and vital state of human physiology [1]. Sleep is essential to life, and is thought to be restorative, conservative, adaptive, thermoregulatory and have memory consolidative functions [2]. Unfortunately, sleep deprivation is placed among the most common stressors experienced during critical illness [3]. In intensive care unit (ICU) clinical practice it is assumed that sleep is important in the recovery process of the critically ill ICU patients and there are strong indications that ICU delirium and sleep deprivation are closely intertwined [4, 5]. In critically ill patients, disturbance of sleep is very common but poorly understood. Polysomnographic (PSG) studies in both mechanically ventilated and non-ventilated critical care patients demonstrate that these sleep disturbances are characterized by severe fragmentation by frequent arousals and awakenings [6, 7]. Sleep architecture is disrupted with a dominance of stage-1 and stage-2 non rapid eye movement (NREM) sleep with reduced deeper phases of sleep (Slow wave sleep (SWS) and rapid eye movement (REM) sleep). For patients in the ICU, sleep traverses the day-night interface, with approximately half of the total sleep time occurring during the day. Total sleep time averages between 2.1 and 8.8 hours of fragmented sleep [8–11]. Spectral composition of the electroencephalogram (EEG) varies as the brain transitions from one sleep stage to the next and each sleep stage has its unique spectral composition. Originally these sleep stages were defined by their unique spectral composition and physiological relevance. With the aberrant EEG states observed in ICU patients, classification is often hampered by the rules of visual analysis, which have to our knowledge never been validated in the critically ill. Drouot et al. recently reported that certain brain states could not be classified according to Rechtschaffen & Kales (R&K) criteria [12], advocating inclusion of two new alternative states for ICU sleep research. The finding of contradictory EEG and electromyography (EMG) activity has been reported previously in other ICU sleep research, often finding sleep-like delta activity in otherwise alert and communicative patients [8, 12, 13]. Very recently Watson et al. reported 85% of all sleep observed in 37 mechanically ventilated ICU patients was of an atypical nature [14]. The increasing interest in the beneficial effects of sleep and the detrimental effects of disturbed sleep has led to a call for automated sleep analysis since manually scoring sleep is both expensive and time-consuming and requires trained personnel [15, 16]. Fortunately, sleep can be analysed, as is increasingly the case, by studying objective properties of the EEG avoiding subjective interpretation [17, 18]. We introduce a novel method to determine depth of sleep from a single EEG channel. It estimates the most relevant aspect of sleep, i.e. depth of sleep over time, and can be used to determine quality and quantity of sleep, specifically in ICU sleep research. We call this method ‘ICU Depth Of Sleep’, or IDOS. The face-validity and physiological basis of

Chapter 3

the new IDOS index is illustrated by comparing application in healthy individuals with R&K visual analysis. This required manually classifying the IDOS index into discrete sleep stages. The method was also compared to R&K classification in a small sample of ICU patients as a proof of concept.

Hypnogram

A Wake REM N1 N2 N3 N4 12:00

15:00

18:00

21:00

00:00

03:00

06:00

09:00

12:00

03:00

06:00

09:00

12:00

IDOS

−1

−2

12:00

15:00

18:00

21:00

00:00

Time of day (hours:minutes)

Figure 1. Hypnogram and IDOS index of the same recording. To illustrate the resemblance of the visually scored R&K (A) hypnogram and IDOS index (B), an example is shown. This recording shows normal sleep and is one of the 15 datasets used for further comparison. The hypnogram resulting from R&K analysis (A) shows a long period of wake EEG with occasional transitional sleep. Towards midnight the EEG transitions to increasingly deep stages of sleep, before gradually resurfacing to shallow stages of sleep. This process is repeated roughly four times, known as ultradian rhythm. The IDOS index (B) of the same recording shows similar transitions of depth of sleep. Simple linear thresholds are manually selected to define the transition from wake (blue) to sleep (red) and the transition to SWS (green). The same method of classification has been applied to all 15 non-ICU recordings and all five ICU recordings

Intensive care unit depth of sleep Proof of concept of a simple electroencephalography index in the non-sedated

Methods IDOS index

Delta power activity is known to decline during the night and to increase as sleep deepens during individual sleep cycles. Conversely, high frequency activity increases as the brain shifts towards wakefulness, particularly in the gamma frequency band [20]. This global property of the EEG during sleep has led to the use of power ratios in sleep analysis, mainly between low frequency and high frequency bands [15, 19, 20]. The gamma and delta band powers for the individual stages of sleep as defined by R&K for a representative recording of normal sleep are given in Figure 1. Dividing gamma power by delta power for individual epochs results in a temporal estimate of depth of sleep, the IDOS index, visually similar to the hypnogram gained by R&K analysis. This visual resemblance is illustrated in Figure 2, where the index is calculated for the same recording as the one used in Figure 1.

Normalized PSD

Gamma

0.2 0.18

0.9

0.16

0.8

0.14

0.7

0.12

0.6

0.1

0.5

0.08

0.4

0.06

0.3

0.04

0.2

0.02

0.1

REM Wake

Delta

REM Wake

Figure 2. The power spectral densities for the R&K stages from the same recording used for Figure 1 (one of the 15 recordings of normal sleep) show the unique spectral composition of different sleep stages and the wake EEG. The central mark of the boxplots represents the median value; the boxes extend to the 25th and 75th percentile. The confidence intervals extend to a maximum of ±2.7 SD. All points outside this range are displayed as outliers, as red plusses. As sleep deepens (towards N4) delta PSD increases at the cost of gamma PSD, with the exception of REM sleep

Chapter 3

A bandpass-filter was applied to remove high-frequency noise and low-frequency drifts and artefacts (caused by breathing, sweating) using a 16th order Butterworth filter between 0.5 Hz and 48 Hz on the single channel EEG data. To stay true to the simplicity and real-time performance of the method, no manual techniques of artefact removal were applied. Discrete short-time Fourier transformation was applied using a 2 second Hamming window with 50% overlap, resulting in a frequency resolution of 0.5 Hz. Spectral densities were then smoothed using a 240 second moving average square window. The delta (0.5-4 Hz), theta (4-7 Hz), alpha (7-12 Hz), beta (12-30 Hz) and gamma (30-48 Hz) powers were obtained by combination of the power spectral densities of the corresponding 0.5 Hz bins. The power in each frequency band was normalized by calculating the power in each frequency band relative to total power in the range 0.5-48 Hz.

Patients and controls Five patients admitted to the ICU of the department of Critical Care of the University Medical Center Groningen, Groningen, The Netherlands were enrolled in our study. Informed consent was obtained from each patient. The local medical ethics committee (Medical Ethical Committee of the University Medical Center Groningen (METc UMCG), research project number 2012.185) reviewed and approved the study protocols. Patients received all aspects of normal care during ICU stay according to standardized protocols. PSG recordings of the controls were obtained from our outpatient clinical database. Fifteen recordings were evaluated as exhibiting sleep without relevant abnormalities and were selected for further use. These patients were referred to the sleep laboratory for suspected sleep apnoea (12, 80%), restless legs syndrome (1, 6.66%), insomnia (1, 6.66%) or chronic fatigue syndrome (1, 6.66%).

Polysomnography Polysomnography (PSG) sleep recording included a six channel (F3, A1, A2, C3, C4, O1) electroencephalogram (EEG), two channel electro-oculogram (EOG) of ocular movements and an electromyogram (EMG) of the left and right masseter muscle or the submental muscles. Furthermore, pulse oximetry and a 12-lead electrocardiography (ECG) were performed. EEG-electrodes were placed according to the international 10-20 system with Ag/AgCl electrodes, sharing the same reference. EEG, EMG, EOG and ECG were sampled at 256 Hz using either an Embla® A10 (Medcare, Reykjavik, Iceland) or Morpheus® (Micromed, Mogliano Veneto, Italy) digital recorder. Patients' skin was prepared according to standard techniques. In controls, additional polygraphic sensors were placed depending on the clinical question e.g. tibial muscle EMG electrodes to detect restless legs. These additional sensors were not used in the ICU recordings. All recordings in the control group were done in an ambulatory setting starting between 10am and 5pm and ending between 10am and 5pm the following day. All recordings were visually scored by the same clinical neurophysiologist using standard R&K criteria, in 30s epochs [21]. The temporal classification of sleep and wake stages was done by visual interpretation of individual epochs in the software environment Brain RT (OSG, Rumst, Belgium). Amongst other PSG derived data, total sleep time (TST), sleep efficiency and percentage spent in

Intensive care unit depth of sleep Proof of concept of a simple electroencephalography index in the non-sedated

each sleep stage were determined to describe quality and quantity of sleep. PSG was performed for a minimum of 24 hours, up to 72 hours depending on patient’s tolerance and ICU length of stay.

Data acquisition After visual analysis, all subsequent analysis was performed using Matlab with the signal processing toolbox (Matlab 2012b, Natick, Massachusetts, United States). After detailed analysis of each recording, a single channel was selected for further analysis. This channel is derived by calculating the difference between C3 and C4, both placed centrally on the left and right hemisphere, known as C3-C4. This electrode location and configuration has been shown to be most representative in distinguishing between relevant sleep-states in healthy individuals with minimal EMG interference [22]. Single channel EEG has the added advantage of low complexity and therefore added speed of clinical setup for possible future application. The all-or-nothing functionality that comes with it reduces the risk of undetected electrode malfunction in future real-time analyses.

Validation The IDOS index was calculated for all outpatient and ICU recordings. Each day of ICU recording was treated as an individual recording during analysis. Although the purpose of the IDOS index is merely to display depth of sleep over time, and not to classify sleep into discrete stages, a semi-automatic comparison to R&K was performed to quantify face validity. To facilitate epoch-by-epoch comparison between the IDOS index and R&K classification, the index was averaged over 30 second segments, followed by manual threshold selection for the transition between sleep and wake with a-priori knowledge of R&K classification for each entire day of recording. A single value of the IDOS index was selected that best resembled sleep onset and offset for each individual day of recording. The same was done for the transition from sleep to SWS. The resulting classifications into two (sleep and wake) and three (wake, non-SWS and SWS) classes were compared to R&K analysis. For purposes of meaningful comparison, the R&K classification was reduced to the same two and three classes by combining NREM stage 1, NREM stage 2 and REM sleep to form non-SWS. SWS consisted of NREM stage 3 and NREM stage 4 sleep. Cohen’s Kappa statistic was used to evaluate agreement between both methods.

Chapter 3

Table 1. Patient characteristics for the control group and results of R&K analysis (n = 15) Characteristics

Control group, n = 15

Male/Female, n

8/7

Age, mean (SD), y

42.9 (16.2)

BMI, mean (SD)

28.8 (9.3)

ESS score, mean (SD)

8.1 (4)

Average duration of PSG, h

19.9 (1.5)

TSTb, mean (SD), hours:minutes

7:50 (1:02)

Sleep latency, means (SD), minutes Time spent in each stage, mean (SD), % of TST

23 (10) b

REM

21 (4)

11 (6)

46 (9)

12 (8)

10 (5)

Sleep efficiency , mean (SD), % c

93 (5)

The Epworth Sleepiness Scale (ESS) is used to determine the level of daytime sleepiness. A score of ten or more is considered sleepy, 18 or more is very sleepy. a

Total sleep time (TST) is the combined time spent in either sleep stage. b

Sleep efficiency is the percentage of time spent asleep relative to the time spent in bed. c

Intensive care unit depth of sleep Proof of concept of a simple electroencephalography index in the non-sedated

Table 2. Contingency table of the pooled results for outpatient recordings (n = 15) The contingency table of visually scored polysomnographic classification (‘R&K’) versus the classification by IDOS index for the combined epochs of all 15 datasets is shown. The number of epochs classified as either wake, non-SWS or SWS are given for both methods. Rows represent the number of epochs scored as each state by conventional visual analysis according to R&K. Columns represent the classification of the same epochs by the IDOS index. The agreement for three classes was excellent (Cohen’s kappa coefficient = 0.82, SD = 0.06), agreement for two classes was slightly better (kappa = 0.84, SD = 0.09) IDOS Wake Wake R&K

non-SWS SWS

non-SWS

SWS

19734

1429

1374

9153

874

653

2571

Chapter 3

Results A total of 298.40 hours of control PSG recordings were obtained from 15 outpatients, with a mean (SD) recording time per patient of 19.89 (1.48) hours (Table 1). The average agreement for these recordings as defined by Cohen’s Kappa for three stage classification was 0.82 (0.06) (Table 2A). For two stage classification Kappa was slightly higher, at 0.84, with a standard deviation of 0.08. The inclusion of five ICU patients yielded approximately nine days of PSG recording, or 205.38 hours. Patient characteristics are summarized in Table 3. Patient A was admitted to the ICU for plasmapheresis, as main treatment for thrombotic thrombocytopenic purpura (TTP)/haemolytic uremic syndrome (HUS). Patient B was intubated and sedated with propofol on the third day of recording. Due to the dominant propofol induced EEG activity, this day of recording was excluded from further analysis. Patient C was included in this study while on mechanical ventilation, and was extubated on the second day. Patient D, with progressive neuromuscular disease, was admitted for non-invasive mechanical ventilation. Patient E was admitted for treatment of an arterial thrombus. Per day manual threshold selection of the IDOS index resulted in similar amounts of the individual sleep stages (not shown) with a high variance of Cohen’s Kappa statistic between recordings. The agreement between both classification methods was excellent for patient A (Kappa = 0.90), reasonable for patient B (Kappa = 0.46), good for patient C (Kappa = 0.65), and excellent for patients D (Kappa = 0.83) and E (Kappa = 0.80). Average sensitivity (SD) for wake epochs was 0.88 (0.10), 0.66 (0.15) for non-SWS, and 0.68 (0.22) for SWS. Average specificity (SD) was 0.87 (0.09) for wake; 0.69 (0.14) for sleep and 0.59 (0.27) for SWS. The classifications of ICU recordings by R&K and the IDOS index are explained in greater detail in Figure 3.

Intensive care unit depth of sleep Proof of concept of a simple electroencephalography index in the non-sedated

Table 3. ICU patient characteristics Patient

Sex

BMI

25-30

20-25

30-35

20-25

>35

Age range, years

60-70

50-60

60-70

40-50

Days prior on ICU

1.81

2.98

2.94

0.99

0.93

Duration of inclusion, days APACHE IIa APACHE IV

TISS 76, meanc Diagnosis on admittance

Mechanical ventilation, days Length of ICU stay, days

15.5

8.5

17.6

Thrombotic thrombocy topenic purpura 0

Bacterial pneumonia

Viral pneumonia

Bacterial pneumonia

0.3

0.5

0.3

1.8

1.3

24.3

21.3

Arterial thrombus

Medication, per recording day, median Benzodiazepines , mg (Lorazepam eqv.) Opioids, mg (morphine eqv.) Propofol 2%, ml

Acute Physiology and Chronic Health Evaluation (APACHE) II is a severity of illness scoring system, with scores ranging from 0 (best) to 71 (worst). All scores were determined using the most abnormal values in the first 24 hours of ICU admission. a

APACHE IV is a severity of illness scoring system, with scores ranging from 0 (best) to 150 (worst). All scores were determined using the most abnormal values in the first 24 hours of ICU admission. b

Therapeutic Intervention Scoring System (TISS 76) quantifies nursing workload into four classes depending on 76 points of therapeutic intervention. The mean is calculated from daily scores. c

Chapter 3

Wake REM N1 N2 N3 N4

IDOS

−1

−2

12:00

00:00

06:00

12:00

18:00

00:00

06:00

12:00

18:00

00:00

06:00

12:00

18:00

00:00

06:00

12:00

18:00

00:00

06:00

12:00

18:00

00:00

06:00

12:00

18:00

00:00

06:00

12:00

18:00

00:00

06:00

12:00

Wake REM N1 N2 N3 N4

IDOS

18:00

−1

−2

12:00

Wake REM N1 N2 N3 N4

IDOS

−1

−2

12:00

Wake REM N1 N2 N3 N4 0

−1

−2

−1

−2

12:00

Wake REM N1 N2 N3 N4

IDOS

10 18:00

00:00

06:00

12:00

Time of day (hours:minutes)

18:00

06:00

Time of day (hours:minutes)

Figure 3. Hypnogram after R&K classification of ICU patients with corresponding IDOS tracings. (Continued on next page)

Intensive care unit depth of sleep Proof of concept of a simple electroencephalography index in the non-sedated

The IDOS index, calculated from EEG channel C3/C4 is shown below the hypnograms after R&K analysis for all five ICU patients. Both R&K classification and the IDOS index of PSG recordings show relatively normal sleep for Patient A, particularly on the second night. The first night shows some arousals, but there is still transitioning to SWS. Nearly all sleep is seen during the night. Patient B shows severe fragmentation, with sleep spread evenly between day and night. In the Hypnogram and IDOS tracing rapid switching between sleep and wake is visible. Patient C’s hypnogram shows a period of sleep in the first night, when the patient was non-sedated and mechanically ventilated. After extubation following the first night, sleep architecture seems to gradually deteriorate. Sleep on days two and three is fragmented with little SWS and no distinct sleep cycles in either tracing. Patient D shows severe fragmentation by awakenings and a period of daytime-sleep, visible in the hypnogram and IDOS tracing. Patient E shows a single ultradian sleep cycle around 2:00 in both tracings, followed by a period of fragmented sleep. All sleep occurred during the evening and night according to the hypnogram, although the IDOS tracing suggests short bursts of sleep during the day

Chapter 3

Discussion Using single channel EEG data, the IDOS index seems to be a promising and simple estimate of depth of sleep, even in the critically ill. Eliminating the need for human intervention in the analysis of the ICU acquired data results in fast, cost-effective and objective insight in ICU patients’ sleep. This opens up possibilities, not only for future large scale ICU sleep research, but also for the monitoring of individual ICU patients for targeted therapy to facilitate natural sleep. Calculating a simple ratio of gamma and delta frequency activity using only a single channel of EEG however has, to our knowledge, never been attempted as an index for depth of sleep. So far, the study of sleep and wake in the ICU has relied heavily on PSG, a timeconsuming and complex method of measuring several parameters most indicative of quality and quantity of sleep [23]. For most patients these recordings take place during the night, however in the ICU sleep is often not limited to the night alone and 24-hour PSG is preferred. The acquired data is manually scored in 30-second segments, which eventually amounts to large workloads and high costs. The R&K classification is for all intents and purposes relatively subjective and inter-rater agreement between individual scorers is consequently low in ICU recordings. Previous studies reporting inter-rater reliability for R&K analysis of ICU recordings show Kappa’s ranging from 0.56 to as low as 0.19 [24, 25]. Using data from only a single EEG channel in a more promising and objective manner, similar results were achieved, without the specific need for other nonencephalographic signals. Although this method inherits some of the disadvantages of PSG, i.e. the dependency on clean electrical activity in a high-tech environment and high variability of ICU patients EEG spectral composition, it simplifies and objectifies the practical aspects of depth of sleep measurement. Attempting to categorize the EEG of sedated patients, or patients with significant neurologic disorders, into discrete stages of sleep seems ineffective. We therefore propose a more objective perspective on EEG activity, while still capable of showing temporal changes in depth of sleep as they occur in ICU patients. To illustrate face validity with the current standard for depth of sleep, R&K analysis of PSG recordings, the IDOS index was manually classified into discrete stages. This method seems justifiable in outpatient recordings, were R&K is widely used. The comparison of IDOS to R&K in ICU patients is less obvious however, since R&K inter-rater agreements are known to be low in these recordings. For future validation in a larger sample of critically ill patients, the correlation of the IDOS index with behavioural assessment, sedation scores, severity of illness and automated methods such as SEF95 and BIS needs to be determined. The exact factors attributable to poor sleep quality in the ICU, and their contributions in disturbing sleep are not yet known. Acute illness, patient-care interactions, light, pain, patient discomfort and noise are all factors that likely contribute to the frequent arousals and awakenings ICU patient’s experience [8, 9, 13, 23, 26–28]. Also, sleep medication that is given to overcome these observed disturbances may result in a state that subjectively

Intensive care unit depth of sleep Proof of concept of a simple electroencephalography index in the non-sedated

resembles sleep, but may not be as physically beneficial as true slow-wave sleep [29]. To increase our understanding of the intricacies of sleep in the ICU, the effects of these potentially disturbing factors need to be studied closely in future studies. This requires application of new techniques, well suited for large scale application in the ICU, one of which could be the IDOS index. The main disadvantage of this study was the limited number of recordings to investigate the practical considerations of the proposed technique in the ICU environment. Severity of illness scores varied significantly between patients, as did other patient characteristics such as days prior on ICU and administered doses of benzodiazepines and opioids. Although results do not justify immediate use in ICU sleep research, they do warrant further development and testing in a more representative cross-section of the general ICU population. From a technical standpoint several limitations of the study should be mentioned. One of the main technical disadvantages of spectral analysis of EEG data is the large interindividual difference between recordings and patients. Despite attempts to minimize these differences by filtering and normalization, there are still visible changes from one day of recording to the next for the same ICU patient, and also between individual patients. Thresholds between days varied in the ICU recordings, and thresholds between patients varied significantly in both groups. This made classification difficult in the most abnormal recordings in the ICU, and hampered agreement with R&K, but does not necessarily diminish the value as an indicator for depth of sleep. The choice to involve gamma-band electrical activity in the analysis of sleep is controversial. The possibility that EMG is responsible for the majority of electrical activity in this range does however not negate the fact that it usable as a variable for sleep state analysis in most non-sedated patients. More importantly, potential noise or artefacts from nursing care activities or other sources could be most apparent in the gamma-band and therefore heavily skew the IDOS index. Minimizing the effects of these changes on the parameters used to determine depth of sleep has first priority in further development of this technique. The study of sleep in the ICU is a growing field of interest. Patients barely seem to sleep for prolonged period of time, if at all, and all criteria for the diagnosis of delirium may be caused by loss of sleep [23]. Our perception of sleep and its relevance in ICU patients’ wellbeing has changed thanks to the introduction of small-scale ICU sleep research relying heavily on PSG. Before large scale interventional studies can be undertaken effectively and efficiently, unsupervised, simple, robust and preferably real-time analysis of sleep is needed. With the IDOS index the first step has been made, using only simple existing techniques, towards scalable ICU depth of sleep monitoring.

Chapter 3

Conclusion Our IDOS index showed excellent agreement with traditional R&K analysis of recordings exhibiting normal sleep, for two and three stage classifications. This indicates solid performance of the index in measuring depth of sleep. The high face-validity in the control group is also reflected in ICU patients with relatively normal sleep. Although agreement between both methods was highly variable in ICU patients, the average agreement seems promising for future clinical and research application. Overall, we conclude that using the new IDOS index, depth of sleep can be determined reliably using only single channel EEG data from outpatient recordings. Future efforts will focus on validating and fine-tuning the index to be used in large scale ICU sleep research.

Intensive care unit depth of sleep Proof of concept of a simple electroencephalography index in the non-sedated

References 1.

2. 3.

10.

11.

12.

13.

14.

15.

16.

Weinhouse GL, Schwab RJ, Watson PL, et al. Bench-to-bedside review: Delirium in ICU patients - importance of sleep deprivation. Crit Care. 2009;13:234. Chokroverty S. Overview of sleep & sleep disorders. Indian J Med Res. 2010;131:126–140. Watson PL. Measuring sleep in critically ill patients: beware the pitfalls. Crit Care. 2007;11:159. Boyko Y, Ørding H, Jennum P. Sleep disturbances in critically ill patients in ICU: How much do we know? Acta Anaesthesiol Scand. 2012;56:950–958. Bellapart J, Boots R. Potential use of melatonin in sleep and delirium in the critically ill. Br J Anaesth. 2012;108:572–580. Bourne RS, Minelli C, Mills GH, Kandler R. Clinical review: Sleep measurement in critical care patients: research and clinical implications. Crit Care. 2007;11:226. Parthasarathy S, Tobin MJ. Applied Physiology in Intensive Care Medicine. Springer Berlin Heidelberg, Berlin, Heidelberg. Cooper AB, Thornley KS, Young GB, et al. Sleep in critically ill patients requiring mechanical ventilation. Chest. 2000;117:809–818. Freedman NS, Kotzer N, Schwab RJ. Patient perception of sleep quality and etiology of sleep disruption in the intensive care unit. Am J Respir Crit Care Med. 1999;159:1155–1162. Edéll-Gustafsson UM, Hetta JE, Arén CB. Sleep and quality of life assessment in patients undergoing coronary artery bypass grafting. J Adv Nurs. 1999;29:1213–1220. Friese RS, Diaz-Arrastia R, McBride D, et al. Quantity and quality of sleep in the surgical intensive care unit: are our patients sleeping? J Trauma. 2007;63:1210–1214. Drouot X, Roche-Campo F, Thille AW, et al. A new classification for sleep analysis in critically ill patients. Sleep Med. 2012;13:7–14. Freedman NS, Gazendam J, Levan L, et al. Abnormal sleep/wake cycles and the effect of environmental noise on sleep disruption in the intensive care unit. Am J Respir Crit Care Med. 2001;163:451–7. Watson PL, Pandharipande P, Gehlbach BK, et al. Atypical sleep in ventilated patients: Empirical electroencephalography findings and the path toward revised ICU sleep scoring criteria. Crit Care Med. 2013;41:1958–1967. Low PS, Shank SS, Sejnowski TJ, Margoliash D. Mammalian-like features of sleep structure in zebra finches. Proc Natl Acad Sci USA. 2008;105:9081–9086. Gervasoni D. Global Forebrain Dynamics Predict Rat Behavioral States and Their Transitions. J

Neurosci. 2004;24:11137–11147. 17. Nieuwenhuijs D, Coleman EL, Douglas NJ, et al. Bispectral index values and spectral edge frequency at different stages of physiologic sleep. Anesth Analg. 2002;94:125–129. 18. Gehlbach BK, Chapotot F, Leproult R, et al. Temporal disorganization of circadian rhythmicity and sleep-wake regulation in mechanically ventilated patients receiving continuous intravenous sedation. Sleep. 2012;35:1105–1114. 19. Brignol A, Al-ani T, Drouot X. Phase space and power spectral approaches for EEG-based automatic sleep-wake classification in humans: A comparative study using short and standard epoch lengths. Comput Methods Programs Biomed. 2013;109:227–238. 20. Stephenson R, Caron AM, Cassel DB, Kostela JC. Automated analysis of sleep-wake state in rats. J Neurosci Methods. 2009;184:263–274. 21. Kales A, Rechtschaffen A. A Manual of Standardized Terminology, Techniques and Scoring System for Sleep Stages of Human Subjects. U.S. National Institute of Neurological Diseases and Blindness, Neurological Information Network, Bethesda, Md. 22. Berthomier C, Drouot X, Herman-Stoïca M, et al. Automatic analysis of single-channel sleep EEG: validation in healthy individuals. Sleep. 2007;30:1587–1595. 23. Weinhouse GL, Schwab RJ. Sleep in the critically ill patient. Sleep. 2006;29:707–716. 24. Elliott R, McKinley S, Cistulli P, Fien M. Characterisation of sleep in intensive care using 24-hour polysomnography: an observational study. Crit Care. 2013;17:R46. 25. Ambrogio C, Koebnick J, Quan SF, et al. Assessment of sleep in ventilator-supported critically III patients. Sleep. 2008;31:1559–1568. 26. Gabor JY, Cooper AB, Crombach SA, et al. Contribution of the intensive care unit environment to sleep disruption in mechanically ventilated patients and healthy subjects. Am J Respir Crit Care Med. 2003;167:708–715. 27. Tembo AC, Parker V. Factors that impact on sleep in intensive care patients. Intensive Crit Care Nurs. 2009;25:314–322. 28. Castro RA, Angus DC, Hong SY, et al. Light and the outcome of the critically ill: an observational cohort study. Crit Care. 2012;16:R132. 29. Zhong X, Hilton HJ, Gates GJ, et al. Increased sympathetic and decreased parasympathetic cardiovascular modulation in normal humans with acute sleep deprivation. J Appl Physiol. 2005;98:2024–2032.

Chapter 4 Automated versus manual scoring of ICU sleep data Submitted for publication Laurens Reinke, Esther M. van der Heide, Pedro Fonseca, Peter Anderer, Anthony R. Absalom, Jaap E. Tulleken

Chapter 4

Abstract Purpose Severe sleep disruption is common among intensive care unit (ICU) patients. Current efforts to better understand and alleviate ICU sleep disruption are limited by the costs of human expert sleep scoring. Automated scoring has practical advantages, but is not validated for ICU sleep recordings. We investigated the performance of a commercially available sleep scoring system, relative to the agreement between two human scorers.

Methods A human expert (M1) and the Somnolyzer 24x7 automated sleep scoring system (A) scored sleep recordings of 61 non-sedated ICU adult patients with durations between 24 and 72 hours. We randomly selected 17 recordings to be scored by an additional human expert (M2). Agreements between the automated system and M1, and between M1 and M2 were calculated and compared. Finally, we investigated the correlation between inter-rater agreements and disease severity and mortality.

Results The agreement between the automated system and M1 (Cohen’s kappa (A∩M1) = 0.40 ± 0.20) did not differ significantly from agreement between human scorers (Cohen’s kappa (M1∩M2) = 0.31 ± 0.18) for paired samples. Disease severity correlated negatively with interrater agreement.

Conclusion The Somnolyzer 24x7 can compete with human expert scoring for ICU sleep recordings. Increased disease severity is associated with worse agreement between human scorers, and between human and automated scorings. This suggests that the electroencephalograms of critically ill patients exhibit patterns that are not easily classifiable according to the AASM criteria for scoring sleep.

Automated versus manual scoring of ICU sleep data

Introduction Sleep is a dynamic, complex physiological process essential for homeostasis, recovery, and survival [1, 2]. Disrupted or delayed sleep is associated with impaired immune function [3], increased susceptibility to infections and impaired wound healing [4, 5], impaired metabolic and endocrine function [6], increased pain perception [7, 8], and impairment of neurophysiologic organization and memory consolidation [9]. Sleep deprivation affects up to 60% of all critically ill patients admitted to an intensive care unit (ICU) [10, 11]. Sleep among these patients is often fragmented by frequent arousals and awakenings which hamper transitions to deeper stages of sleep, reduced duration of sleep, and disturbed distribution of sleep with up to half of the total sleep time occurring during the day [4, 5, 11, 12]. Poor sleep during critical illness is considered to be a major stressor for patients during and after ICU admission. It is associated with the development of ICU delirium and long-term cognitive decline, and has detrimental effects on recovery, morbidity, and mortality [13–15]. The ICU is a unique environment where a multitude of intrinsic and environmental factors may hamper sleep [16–22]. Although previous studies have provided new insights into the aetiology and possible prevention of disturbed sleep in the ICU, their scope, statistical significance and reliability have thus far been constrained by the logistical challenges of measuring and assessing sleep objectively [2, 4, 20, 22–28]. Electroencephalography (EEG) has historically been the primary tool for objective sleep monitoring [29, 30]. When EEG is combined with electromyography (EMG), and electrooculography (EOG), to investigate sleep, the technique is known as polysomnography (PSG). The visual and manual annotation or scoring of these recordings commonly follows criteria originally set by Rechtschaffen and Kales [31], with additional changes later culminating in the American Academy of Sleep Medicine (AASM) Manual for the Scoring of Sleep [32]. Hundreds or even thousands of 30 second epochs each comprising multiple channels of PSG data are typically processed by a single human expert. Although this method is considered to be the gold standard for routine clinical sleep analysis, most PSG studies in critically ill patients report difficulties in setting up, maintaining, and manually processing and scoring ICU sleep recordings [4, 12, 33–36]. The practical expertise required to apply and maintain the array of electrodes required for human scoring further limits scalability and increases costs. Furthermore, the reliability and repeatability of manual analysis of ICU sleep recordings is lower than for other clinical recordings [37]. While Elliott et al. reported observed ‘reasonable’ to ‘good’ agreement between two combinations of three human scorers in discerning wake from sleep activity, the agreement on detailed sleep staging was much lower depending on individual sleep stages and the combination of human scorers [23]. The rapid development of automated systems seems to be a step forward, promising reproducible and traceable scorings, from limited data, without costly human

Chapter 4

intervention. Although commercially available for clinical application, no system capable of full AASM sleep scoring has been fully evaluated for use in the ICU. We therefore investigated the agreement between the scores produced by a commercially available AASM scoring system (Somnolyzer 24x7) and those of a human scorer, relative to the agreement between two human scorers. We hypothesized that automated scoring would reach agreement with that of a human scorer, comparable to that between two human scorers, while providing practical advantages over the current gold standard for sleep analysis.

Automated versus manual scoring of ICU sleep data

Materials and Methods Study population and patient recruitment We obtained 70 PSG recordings during an observational study (Trial NL5197, NTR5345) primarily investigating the influence of disrupted biorhythms on the quantity and quality of sleep among patients of the department of Critical Care of the University Medical Center Groningen (UMCG). After approval by the local ethics committee (UMCG METc, registration number 2015/00295), data collection started in September 2015 and finished in September 2018. All adult patients without a history of sleep pathology, an expected ICU stay of at least 48 hours, and a Richmond Agitation and Sedation Scale (RASS) above -3 were eligible for inclusion in the study. Informed consent was gained from patients with capacity to do so. For patient lacking capacity, informed consent was first obtained from their legal representatives, followed by consent after they recovered consciousness. Neurosurgical patients, and patients taking melatonin supplements were excluded from participation. The sample size needed to quantify the performance of the automated system relative to the inter-rater agreement between human scorers could not be estimated a-priori due to the severity and uneven distribution of sleep disruption among ICU patients, the unpredictable degree of class imbalance between individual sleep stages, and the high variation of inter-rater agreement between human experts [23, 37, 38].

Data acquisition The following data were acquired from the digital patient record: age, BMI, sex, reason for ICU admission, length of hospital and ICU stay, 12-month mortality, specific daily medication and sedative use, and treatment with mechanical ventilation. The Acute Physiology and Chronic Health Evaluation 4 (APACHE IV) and the Simplified Acute Physiology Score 2 (SAPS II) were calculated on the day of ICU admission.

Polysomnography PSG was recorded for a period of 24-72 hours depending on patient’s tolerance, RASS scores, and ICU length of stay. The recording consisted of six EEG channels (F3, A1, A2, C3, C4, O1), two EOG channels and EMG of the left and right masseter or submental muscles. Ag/AgCl electrodes were used, with EEG-electrodes being placed according to the international 10-20 system after skin preparation according to standardized techniques. Signals were amplified using a BrainAmp DC32 amplifier with a BrainVision recorder (Brain Vision Solutions, Montreal, Canada) and an Alice 6 LDx system (Philips Respironics, Murrysville, USA). Anonymized data were then stored at 256 Hz for subsequent analysis by two experienced human experts and by the automated system.

Sleep analysis All recorded data were blindly assessed for data quality by a human expert, and sets of sufficient quality were then analysed by the automated system (A) and a human expert (M1). We randomly selected 20 patients for further analysis by an additional human expert (M2). Human experts and the automated system were free to select either the C4-A1 or

Chapter 4

C3-A2 EEG channel for AASM analysis depending on signal quality. Somnolyzer uses the recommended C4-A1 channel by default and switches automatically to the backup C3-A1 channel in case of significant artefacts in the default channel. The scoring of discrete wake and sleep stages (REM, N1, N2, N3) according to the latest AASM scoring guidelines by the human scorers was done by visual interpretation of individual 30 second epochs in the Brain RT software (OSG, Rumst, Belgium). The Somnolyzer 24x7 scoring system version 3.2 (Philips Respironics, Murrysville, USA) performed the same classification automatically. A detailed description of the Somnolyzer 24x7 automated scoring system was published in 2005 [39] for sleep scoring according to Rechtschaffen and Kales and in 2010 [40] for sleep scoring according to AASM, with further validation studies in non-critically ill subjects published more recently [41]. Uninterpretable and artefact-filled epochs were labelled for exclusion from further analysis by the automated system and human experts. Only epochs and datasets deemed classifiable by all three scorers were used for further statistical analysis. Further spectral analysis of EEG data to investigate possible sources of disagreement was performed using the previously described IDOS-index [42]. This index represents the ratio between activity generally associated with wakefulness and activity increasingly indicative of sleep.

Statistical analysis Sleep-related parameters were calculated using Matlab (Matlab 2014b, Natick, MA, USA). Statistics were calculated using SPSS 24 (2016, IBM, Armonk, NY, USA). The MannWhitney U test was used to test the significance of differences in outcome parameters after stratification (scored by A and M1, versus scored by A, M1, and M2). Cohen’s Kappa statistic was used to evaluate agreement between human expert scorers and the automated system. Cohen’s kappa corrects for chance agreement due to imbalanced datasets, such as the imbalanced distribution of sleep and wake stages. Scoring performance statistics for wake and individual sleep classes were calculated using a binary one-vs-rest strategy. For normative interpretation of inter-rater agreement, we used the guidelines by Landis and Koch [43]. Spearman’s correlation coefficient was used to quantify the correlation between inter-rater agreements, disease severity, and mortality. For estimation of statistical significance, an alpha of 0.05 was used. Unless indicated otherwise, results are presented as mean values (standard deviation).

Automated versus manual scoring of ICU sleep data

Obtained informed consent: 70

Data lost due to technical failure: 4

Data-analysis by the automated system and M1: 66

Randomly selected for additional analysis by M2: 20

Insufficient signal quality: 5

Statistical analysis: 61

Insufficient signal quality: 3

Statistical analysis: 17

Figure 1. Patient and data inclusion flow chart

Chapter 4

Results Seventy patients were included in the main study. PSG data from four (5.71%) were lost due to undetected technical failure of EEG equipment during measurement (Figure 1). A further five (7.14%) recordings were deemed entirely unscorable by the human experts, leaving 339901 30-second epochs (2832.51 hours) for further analysis. Of the 20 recordings selected for classification by a second human scorer, three (15%) were rejected entirely due to low signal quality, leaving 51454 epochs (428.78 hours) for analysis of inter-rater agreement between two human scorers. Patient characteristics are summarized in Table 1. Recordings randomly selected for additional classification by M 2 were from patients with a significantly shorter median ICU stay (14.01 (6.02-29.51 IQR) days versus 7.01 (4.0118.98 IQR) days, P = 0.014), but there were no other significant differences with the rest of the sample (supplementary Table 1). The aggregated confusion matrix for the classification by the automated system (A) and the human scorer (M1) for all 339901 epochs (Figure 2) shows a general over-classification of REM and N1 sleep by the automated system. Additional substantial disagreement is seen between individual sleep classes and the wake class, and between N2 and N3. The aggregated confusion matrix for agreement between the human scorers (M1∩M2) is shown in supplementary Figure 1. The automated system reached moderate agreement with M1 for the wake class (κ = 0.56 ± 0.23), and fair agreement for the N2 (κ = 0.33 ± 0.19) and N3 (κ = 0.33 ± 0.29) classes, as shown in Table 2. Due to a general lack of REM sleep in most recordings (0.08 ± 0.19 hours / 24 hours) and low sensitivity (13.19 ± 25.99%), agreement was low for this class (0.09 ± 0.18). Low sensitivity for N1 sleep (1.19 ± 2.79%) resulted in near-chance agreement between the automated system and M1 for N1 (κ = 0.01 ± 0.04). Similar disagreement was found between M1 and M2, as described in supplementary Table 2. The overall inter-rater agreement of the automated system with a human scorer (κ(A∩M1) = 0.40 ± 0.20) did not differ significantly from the agreement between two human scorers (κ(M1∩M2) = 0.31 ± 0.18) for paired samples (n = 17, P = 0.55). Increasing agreement between automated and human classification correlated strongly with increasing agreement between two human experts (r = 0.661, P = 0.004), see Table 3. Furthermore, for recordings from subjects with a high predicted mortality as calculated by the SAPS II metric, there was lower agreement between automated and human classification (r = -0.266, P = 0.042) as well as lower agreement between the human scorers (r = -0.506, P = 0.038). Additionally, disagreement between the human scorers showed moderate correlation with increased 12-month mortality (r = -0.524, P = 0.031), whereas disagreement between the automated system and M1 did not correlate with 12-month mortality (r = -0.175, P = 0.185). Further spectral and visual analysis using the IDOS-index did not result in uniformly distributed abnormalities that could explain inter-rater disagreement (Fig. 3-5). Our human

Automated versus manual scoring of ICU sleep data

expert scorers reported that only a few of the recordings exhibited EEG patterns that were clearly recognizable as physiological sleep with normal architecture. Figure 3 shows an example of a recording with recognizable physiological sleep architecture and high agreement between the automated system and M1 (κ(A∩M1) = 0.80). Figure 4 illustrates one of multiple recordings where moderate agreement was reached despite abnormal EEG patterns (κ(A∩M1) = 0.57), potentially due to a consistent difference between fast and slow wave activity for wake and deeper stages of sleep. Classification of most recordings without this clear spectral separation between fast and slow wave activity for wake and sleep resulted in only fair agreement, compounded further by the abnormal architecture and severe fragmentation. An example of a representative recording is given in Figure 5 (κ(A∩M1) = 0.25).

Predicted (A)

Wake

True (M 1 )

REM

Wake

REM

83.42%

1.47%

6.16%

7.44%

1.51%

(179555)

(3160)

(13268)

(16015)

(3249)

11.61%

59.69%

9.84%

18.24%

0.62%

(112)

(576)

(95)

(176)

(6)

33.87%

10.3%

32.61%

21.3%

1.92%

(5466)

(1663)

(5263)

(3438)

(310)

26.05%

3.95%

9.03%

45.02%

15.94%

(19448)

(2950)

(6744)

(33608)

(11898)

15.74%

0.37%

1.74%

20.48%

61.67%

(5180)

(122)

(571)

(6737)

(20291)

Figure 2. Confusion matrix for scoring by the automated system (A) versus traditional scoring by human expert (M1). Percentages are calculated from class-totals as scored by M1

Chapter 4

Table 1. Demographics Characteristic

N (%)

Sex, female

23 (37.7)

ICU admission diagnosis surgical non-surgical 12-month mortality

19 (31.15) 42 (68.85) 19 (31.15) Median (IQR)

BMI

26 (23-29)

Age, years

60 (52-67)

Duration of hospital stay, days Duration of ICU stay, days Duration of recording, hours

31 (19-55.49) 10.99 (5.51-25.50) 47.74 (17.22)

APACHE IV

69 (47-82)

SAPS II

40 (31-49)

Mechanical ventilation, days

7 (1-16) Mean (SD)

Medication dose per recording day

Benzodiazepines, mg (Lorazepam eqv.)

1.55 (4.82)

Opioids, mg (morphine eqv.)

1.80 (3.63)

Propofol 2%, ml

1.81 (7.27)

Automated versus manual scoring of ICU sleep data

Table 2. Agreement between automated system (A) and human expert (M 1) in 61 PSG recordings Class

Wake

REM

5 classes

Prevalence (M1), hours per day (SD)

15.32 (6.17)

0.08 (0.19)

1.19 (1.33)

5.25 (3.98)

2.17 (3.88)

Kappa, - (SD)

0.56 (0.23)

0.09 (0.18)

0.01 (0.04)

0.33 (0.19)

0.33 (0.29)

0.40 (0.2)

Accuracy, % (SD)

83.1 (9.26)

94.53 (11.88)

92.46 (5.37)

70.61 (14.7)

89.35 (9.35)

65.02 (16.94)

Sensitivity, % (SD)

69.96 (22.56)

13.19 (25.99)

1.19 (2.79)

77.41 (18.14)

51.85 (35.87)

Specificity, % (SD)

88.97 (10.22)

99.26 (1.31)

99.54 (1.07)

63.82 (23.01)

95.16 (6.7)

PPV, % (SD)

80.64 (22.27)

26.77 (34.9)

17.16 (22.57)

45.92 (16.78)

49.04 (32.45)

Performance statistics for individual classes were calculated using a binary one-vs-rest strategy.

Table 3. Spearman’s correlation coefficient for inter-rater agreements and disease severity κ(A∩M1)

Apache IV p

12-month mortality

SAPS II p

κ(A∩M1)

n = 61

- -0.247

0.074

-.266

0.042

-0.175

0.185

κ(M1∩M2)

n = 17

0.661

0.004 -0.259

0.332

-.506

0.038

-.524

0.031

Chapter Chapter 4 4

10 -2 10 -3 10 -4 10 -5 15:00

Wake

REM

15:00

Wake

REM

15:00

Wake

REM

15:00

18:00

21:00

00:00

03:00

06:00

09:00

12:00

15:00

18:00

21:00

03:00

IDOS index

00:00

03:00

06:00

Human rater (M )

21:00

00:00

03:00

06:00

Human rater (M )

21:00

00:00

03:00

06:00

Automated system (A)

21:00

09:00

12:00

15:00

18:00

21:00

00:00

03:00

06:00

09:00

12:00

Time (hours:minutes) Figure 3. Recording of a 59-year old tracheally ventilated subject. This recording lasted for more than two and a half days Figure 3. Recording ofagreement a 59-year old tracheally lasted for more shows than two and a half days with good inter-rater after analysis ventilated (κ(A∩M1) = subject. 0.79, κ(MThis hypnogram well consolidated 2∩Mrecording 1) = 0.74). The with good inter-rater agreement afterfragmentation analysis (κ(A∩M = 0.79, κ(Mlack ) =REM 0.74).sleep. The hypnogram shows well consolidated 1) a 2∩M1of sleep architecture despite significant and general The human scorers reported that the sleep fragmentation a general lack of REM sleep. The human scorers reported that the AASM architecture sleep scoringdespite criteriasignificant were relatively simple to and apply AASM sleep scoring criteria were relatively simple to apply

72 72

15:00

16:00

17:00

18:00

19:00

20:00

21:00

22:00

23:00

00:00

02:00

03:00

04:00

02:00

03:00

04:00

02:00

03:00