Vital-sign circadian rhythms in patients prior to discharge from an ICU: a retrospective observational analysis of routinely recorded physiological data

Background: Circadian deregulation in patients treated in an intensive care unit (ICU) is commonplace and is associated with complications such as immune system disruption and delirium. The presence and nature of circadian rhythms in the vital signs recorded in the ICU are not well documented, nor is their generalisability across different ICU populations. This paper investigates the presence of circadian rhythms in the 24 h prior to discharge from an ICU of patients who subsequently recovered. We hypothesise that vital-sign circadian rhythms will be observable in this cohort of patients, that these circadian rhythms will resemble known behaviour in healthy individuals, and that these circadian rhythms will be generalisable across different populations of ICU patients. Methods: Circadian rhythms are investigated across several commonly measured vital signs: systolic blood pressure, heart rate, respiratory rate, and temperature. The data employed in this paper are from patients in the MIMIC-III (2001–2012), eICU-CRD (2014–2015), and PICRAM (2009–2015) databases, spanning 198,205 patients across 211 hospitals in the USA and the UK. Evaluation of circadian rhythms encompasses a comparison between the observed rhythm profiles and peak-nadir excursions with those found in the literature, as well as the assessment of the correlation in rhythm profiles between databases. Results: Circadian patterns in all four vital signs were found to conform to those reported for non-ICU cohorts. Additionally, all vital-sign circadian profiles were correlated between databases at the p = 0.05 level. The peak-nadir excursion in the observed rhythms was suppressed by a factor of 2–5 relative to results found in the literature for cohorts of young, healthy individuals. Conclusions: Across three different ICU the 24 h prior to discharge from an ICU. This result has potential uses in monitoring patient recovery and early detection of complications such as delirium.


Introduction
Patient care in an intensive care unit (ICU) typically involves maintaining homeostasis or 'normalisation' of vital signs [1][2][3], where the body is unable to provide this for itself. However, the process of controlling and regulating vital signs, combined with sedation, inflammation, environmental light, and noise levels, can disrupt a patient's natural circadian rhythms [4]. ICU practice in general does not emphasise support of a patient's circadian rhythms, though there is a growing desire to improve upon this [5]. Chronically disrupted circadian rhythms are associated with metabolic disorders such as obesity and diabetes, cardiovascular disease, and cancer [6][7][8][9]. In an ICU, disruption or loss of a patient's circadian rhythms is associated with complications such as immune system disruption [10], delirium [11,12], and mortality [13,14].
The assessment of circadian behaviour in the ICU typically focuses on the study of sleep, ideally recorded using polysomnography [15,16]. However, difficulties with instrumentation in the ICU [17], abnormal electroencephalography (EEG, brain activity) patterns [16,18,19], and the relative sensitivity of sleep to events such as lighting or environmental noise variations mean that sleep is not necessarily an ideal or easily established marker of patient circadian behaviour. Thus, a recent review commented that 'Finding the optimum tool to monitor (circadian rhythms in) critically ill patients therefore remains a key to research progress in this area' [3]. Healthy individuals exhibit circadian rhythms in several vital signs, including systolic blood pressure (SBP), heart rate (HR), respiratory rate (RR), and core body temperature (T) [3]. However, the 'severe circadian deregulation' [4] experienced by patients treated in ICUs can result in abnormal vital-sign patterns.
Several studies have established typical circadian vitalsign behaviour in healthy individuals. Hermida et al. [20] conducted a study using ambulatory monitoring on 278 healthy individuals with mean ± SD age 22.7 ± 3.3 years, synchronising measurements to individual sleep/wake times rather than time of day. They observed an elevated SBP during the day, which reached an approximate plateau (with 3 periodic local maxima) between ≈ 3 and ≈ 13 h after awakening, followed by a sinusoidal dip during sleep. The observed rhythms in HR closely corresponded to those observed in SBP, with an elevated plateau (again with 3 periodic local maxima) between ≈ 2 and ≈ 14 h after awakening, decreasing slightly during the day, and a sinusoidal dip overnight.
Bosco et al. [21] conducted a study in which 6 males (competitive scuba divers with mean ± SEM age 39 ± 3 years) were kept in a constant routine protocol (sustained wakefulness, minimal activity). RR was observed to peak late in the day (≈ 8 pm) with a trough at ≈ 3-7 am, roughly in phase with HR. Spengler et al. [22] conducted a similar study in which 10 healthy males with mean ± SD age 23.7 ± 3.9 years were kept in a relaxed, semi-recumbent position isolated from any indication of time of day for 41 h. As in [20], measurements were synchronised to individual sleep/wake times. While they do not report RR, Spengler et al. report ventilation (V E ) in l/min, which was elevated between ≈ 2 h before awakening and ≈ 8 h after awakening, and decreased to an approximate plateau overnight. Core body temperature showed an approximately sinusoidal form, with nadir that lagged behind the nadir in V E by ≈ 6-8 h.
Given previous work has suggested circadian behaviour is severely disrupted in an ICU [3,4,15,23], and assuming that circadian behaviour in the majority of patients returns to normality post-ICU discharge, patients treated in an ICU undergo a 'circadian recovery' process as part of their overall recovery. If this 'circadian recovery' process was shown to begin prior to ICU discharge in patients who subsequently recovered (i.e. were discharged home), and this circadian state was shown to be generalisable across different ICU populations (i.e. not due to external behaviour such as nursing shift changes), there are several potential clinical applications. These include the monitoring of ICU patient recovery, as well as monitoring the development of complications associated with disrupted circadian rhythms such as delirium. We hypothesise that vital-sign circadian rhythms will be observable in the 24 h prior to discharge from the ICU in patients who subsequently fully recovered, that these circadian rhythms will resemble known behaviour in healthy individuals, and that these circadian rhythms will be generalisable across different populations of ICU patients. We set out to validate these hypotheses across three large, retrospective clinical databases.

Databases
This study makes use of three clinical databases: Information Mart for Intensive Care III  (MIMIC-III)  8. Measurements outside of the broad physiological bounds (60 mmHg < SBP < 280 mmHg, 30 bpm < HR < 240 bpm, 4 breaths/min < RR < 60 breaths/min, 34°C < T < 40°C) were excluded. 9. Measurements taken while the patient was under the effect of treatments that were likely to significantly affect the vital signs being measured were excluded. This process focused on removing measurements taken while vasopressors, β-blockers, or other blood pressure medication were likely to be active, and is discussed in more detail in Additional file 1. Vital-sign measurements were excluded: • Up to 1 h after a patient was administered dobutamine, dopamine, adrenaline/epinephrine, noradrenaline/norepinephrine, metaraminol, glyceryl trinitrate, dopexamine, nitroprusside, or isoprenaline [27,28]. • Up to 2 h after a patient was administered vasopressin, propofol, magnesium sulphate, ephedrine, or phentolamine [29][30][31]. • Up to 24 h after a patient was administered milrinone, terlipressin, labetalol, metoprolol, or hydralazine [27,32,33]. 10. The patient must have had at least one night-time (12 midnight-5:59 am) and one day-time (10 am-7:59 pm) SBP measurement as in [34]. The majority of patients will have more available SBP measurements than this (see Additional file 2), but this requirement ensures each patient contributes to both 'day-time' and 'night-time' behaviour. 11. For MIMIC-III and PICRAM, if the patient had multiple ICU stays within 6 months of each other, all ICU stays within this period were excluded due to it being unlikely the patient was discharged 'healthy'. An ICU stay is defined as a period during which a patient occupied a bed in the ICU, including short-term removal for surgery, scans, or other interventions. A hospital admission is defined as the time between patient admission to and discharge from the hospital. In the eICU-CRD, no relative dates are recorded for hospital admissions. Instead, any hospital admission containing multiple ICU admissions was discarded entirely. 12. For each ICU stay, all measurements in each 1-h period were averaged for each vital sign for the final 24 h of that ICU stay. This process avoids weighting data towards ICU stays where patients are more ill, and thus likely to have more regular vital-sign measurements. These mean hourly values were recorded left aligned (e.g. the mean of measurements between 1:00 am and 1:59 am was recorded as occurring at 1:00 am). Vital signs were typically measured at least hourly, with the exception of temperature in MIMIC-III and PICRAM which was typically measured once every 4 h. If there were no measurements of a given vital sign in a given 1-h period in an ICU stay, that ICU stay did not contribute any measurement for that hour to the overall analysis.

Data analysis
Patients were separated into groups by gender and age, as there are established trends in mean SBP and HR associated with gender and age [34]. Observation of similar trends in the selected ICU cohorts would give support to the notion that underlying physiological, rather than treatment or pathology driven, behaviour is being observed in these patients. The age groups used in this paper are a modified set of those specified in 'Provisional guidelines on standard international age classifications' for 'health, health services and nutrition -morbidity and handicaps' [35]. These age groups are as follows: 15-45 years (combining the recommended 15-25-and 25-45-year groups due to the low number of patients under 45 treated in ICUs), 45-65 years, and 65+ years. Per HIPAA regulations, the ages of individuals greater than 89 were not recorded in MIMIC-III or eICU-CRD. These patients were treated as 91 years of age; thus, all fell within the 65+-year age group. The median Oxford Acute Severity of Illness Score (OASIS), a severity of illness score used for predicting patient outcomes [36], was determined for each patient subgroup. The OASIS was designed to require a minimal set of physiological parameters. In contrast, common severity of illness scores such as APACHE and SAPS employ a wide variety of physiological measurements that are not necessarily well recorded or easy to recover from large ICU databases. As such, OASIS is more easily and consistently applicable across a range of retrospective clinical databases with different recording standards.
Evaluation of circadian rhythmicity was performed using several approaches, which were performed using the 24-h mean vital-sign profiles established for each patient cohort. The observed profiles were visually compared to circadian vital-sign profiles found in the literature, typically available for non-ICU cohorts. As a quantitative indication of rhythm amplitude or strength, the peaknadir excursion [37] was calculated, expressed as both a raw value and as a percentage of the 24-h mean for that vital-sign profile. These values were compared to values reported in the literature. To evaluate the consistency of corresponding vital-sign profiles between databases, the cross correlation (R) and accompanying p values (p) were calculated. For the correlation analysis, temperature profiles from eICU-CRD were subsampled at 4-hourly intervals to allow for comparison with MIMIC-III and PICRAM.
To provide an indication of intra-cohort variability, the hourly 95% confidence intervals (CIs) of the vital-sign mean were calculated for each 24-h vital-sign profile [38]. Graphically, if two vital-sign CIs do not overlap for any given hour, their means are significantly different at the p = 0.05 level. Where comparison between databases is desired, a two-sample Student's t test was used to compare each hourly bin of vital-sign measurements. As before, mean vital-sign levels were deemed significantly different if each of the 24 hourly bins was found to be significantly different at the p = 0.05 level.

Database and cohort demographics
Tables 1 and 2 present demographic data for the entire databases and the selected cohort from each database, respectively. The overall median age of patients in the selected cohort for PICRAM (61.2 years) was greater than that for the corresponding cohort in MIMIC-III (59.6 years) or eICU-CRD (60.0 years). Similarly, the overall median OASIS of patients in the selected cohort PICRAM (33) was greater than that for MIMIC-III (27) or eICU-CRD (26). These results suggest that on average, the selected PICRAM patients were older and more ill, corresponding to the increased LOS observed in the selected PICRAM cohort (Table 2).   (11) 25 (10) 24 (11) 24 (12) 33 (15) 30 (16) 45-64 26 (9) 27 (10) 25 (12) 26 (12) 33 (16) 33 (17)   65+ 28 (9) 29 (9) 27 (11) 29 (11) 33 (18) 34 (18) Overall 27 (10) 27 (10) 26 (12) 27 (12) 33 (17) 33 (16) * Length of stay Additionally, median OASIS were identical for the PICRAM cohort selected in this paper and the overall PICRAM database (33), unlike MIMIC-III (27 selected and 29 overall) and eICU-CRD (26 selected and 28 overall). This suggests the employed selection criteria were less discriminatory in PICRAM. This notion is supported by Table 3, which shows the number of patients (#Pat.), hospital admissions (#Hosp.), ICU stays (#ICU), and vitalsign measurements (#SBP, #HR, #RR, #T) that met the cumulative application of the criteria set out previously for each database. In this table, it can be observed that a higher portion of PICRAM ICU stays were retained (23.5%) by the selection process than for MIMIC-III (19.3%) or eICU-CRD (17.5%). Despite this higher retention rate, the selected cohort of PICRAM ICU stays (3283 ICU stays, Table 2) is still significantly smaller than the size of the selected MIMIC-III (11,872 ICU stays) or eICU-CRD (35,143 ICU stays) cohorts. Figure 1 shows the circadian profiles for SBP, HR, RR, and T grouped by gender for MIMIC-III, eICU-CRD, and PICRAM, with 'night-time' represented from 12 midnight-5:59 am and 'day-time' from 10 am-7:59 pm. By visual inspection, these profiles correspond well to those reported in healthy cohorts [20,22] described previously and to those reported for non-ICU patients [34,39]. In MIMIC-III and eICU-CRD, SBP is elevated between ≈ 2 and 14 h after the end of night-time, with three periodic maxima, though these are more pronounced than those observed in [20]. This elevated period is followed by a sinusoidal dip during night-time. HR in MIMIC-III and eICU-CRD is similarly elevated between ≈ 2 and 14 h after the end of night-time, again with three periodic maxima. Elevated SBP and reduced HR for men relative to women (p < 0.05, Fig. 1) also correspond to observations in [20]. In both SBP and HR, the smaller sample size in PICRAM results in more variability and makes features (especially maxima) more difficult to distinguish. However, the overall periods of elevated and reduced SBP and HR in PICRAM appear similar to those observed in MIMIC-III and eICU-CRD. The PICRAM cohort has an elevated mean HR (p < 0.05) relative to MIMIC-III and eICU-CRD and SBP (p < 0.05) relative to MIMIC-III.

Circadian vital-sign qualitative analyses
There is a resemblance between the profiles in RR observed in Fig. 1 and the profiles in RR and V E reported in [21] and [22], respectively. Elevated RR can be observed between ≈ 2 and 14 h after the end of night-time, peaking at 8 pm, with a trough between ≈ 2 and 6 am. T shows the expected sinusoidal behaviour [22,40], though the low measurement frequency in MIMIC-III and PICRAM makes this harder to discern. T also shows a lag in the nadir of approximately 6-8 h relative to RR (and indeed SBP and HR) as observed relative to V E in [22].
The vital-sign patterns mentioned above largely hold for smaller cohorts grouped by gender and age, as shown in Fig. 2 for men and Fig. 3 for women. Once again, the smaller of these cohorts, such as those from PICRAM or the younger 15-44-year cohorts, show a greater degree  of variability which makes features more difficult to distinguish. Figures 2 and 3 also show expected age-related trends [34]. These trends include progressively decreased HR in older age groups (p < 0.05) for MIMIC-III and eICU-CRD and between the 45-64 and 65+ groups in men in PICRAM. Women also show the expected increase in SBP with age (p < 0.05 for MIMIC-III and eICU-CRD); however, this trend is largely absent in men. A consistent increase in magnitude and duration of ascent prior to the morning (8:00 am) SBP peak across men and women as they age can be observed, similar to the trends reported in [34]. RR and T do not show clear variations with age, but both show consistent 24-h patterns across all age groups. Table 4 shows that the peak-nadir excursions in all three ICU databases were attenuated relative to the peak-nadir excursions reported for non-ICU cohorts in the literature. The peak-nadir excursions for SBP and HR in the ICU cohorts in Table 4 are a factor of 4-5 times smaller than the values reported in [20]. Similarly, the peak-nadir excursions for RR in the ICU cohorts are a factor of 2 smaller than the value reported in [21], and the peak-nadir excursion in temperature is a factor of 2-3 times smaller than the corresponding value in [22]. Table 5 shows that there is a strong correlation in vital-sign trends between all of the three databases. All vital-sign profiles were correlated between databases at the p = 0.05 level, and 20 of 24 correlated at the p = 0.01 level. Of the four exceptions, three were temperature profiles, where the lower p values observed were likely due to the lower frequency of the available measurements (once every 4 h).

Presence of circadian rhythms
From Table 5, we can reasonably assert we are observing a generalisable vital-sign circadian 'rhythm' (i.e. a recurring vital-sign pattern with 24-h periodicity). This assertion is based on the high cross-correlations between 24-h vitalsign profiles from different databases, which are subject to different demographics and standards of care. That each individual's contributing vital-sign profile may begin and end at any point within the 24 h adds further credence to the physiological, rather than environmental, origin of the observed rhythmicity.
Further evidence that we are observing vital-sign circadian rhythms is provided in Figs. 1, 2, and 3, where the observed vital-sign profiles show similar patterns across databases and with respect to those reported in the literature for non-ICU cohorts [20,22,40,41]. Additionally, the relative trends between genders and between age groups are consistent with the literature, and across databases [34]. While a previous study [42] noted variations related to time of day in the agreement between nurse-verified and waveform-derived vital-sign measurements in MIMIC-II, these variations were of a 'clinically insignificant amount' , and only measurement variability, not measurement bias, showed significant variation with time of day. As such, this behaviour is unlikely to contribute significantly to the observed profiles.
Overall, these results suggest observation of an intrinsic, consistent, demographically modified 24-h pattern in vital signs observable in the last 24 h prior to discharge from an ICU in the selected cohort of patients. This behaviour, observable across 50,298 ICU stays drawn from 211 hospitals across the UK and the USA with different patient demographics and standards of care, suggests that there is a typical circadian pattern in vital signs present in patients near recovery and discharge from an ICU.

Rhythm topography
The peak-nadir excursions in SBP, HR, RR, and T (Table 4) were found to be 2-5 times smaller than those reported in the literature for healthy cohorts [20,22]. There are several potential causes for this apparent attenuation of circadian amplitude. The suppression may be due to pathology or medication in the selected ICU cohort. The computation of average rhythms does not distinguish between amplitude attenuation caused by a mix of 'healthy' and attenuated rhythms and a consistent, cohort-wide attenuation, though the narrow 95% CIs of the mean would lend support to the latter. Alternatively,  the observed reduced amplitudes may be demographically driven, as the results in both [20] and [22] are for young, healthy adults, and the results in [21] for competitive scuba divers, as opposed to the more heterogeneous, and generally older, cohort employed in this study. Also of note is that the data in both [20] and [22] are synchronised for waking time, rather than clock time, which may accentuate circadian rhythms. However, one would expect some degree of synchronicity in waking time within a given ICU, and for reduced synchronicity to 'smear' or laterally shift patterns rather than significantly decrease their peak amplitude.
A final potential cause of circadian amplitude attenuation is the fact that patients in an ICU are typically recumbent and physically inactive, which can affect circadian rhythm amplitude [43]. This consideration is supported by the fact that rhythm amplitude showed a factor of 4-5 times reduction in HR and SBP compared to [20], where subjects were ambulatory, but only a reduction of 2-3 times in RR and T compared to [21,22], where patients were recumbent and inactive. Despite the intuitive appeal of these results, caution should be taken as [20][21][22] report different sets of vital signs using different protocols and equipment, and [21,22] contain data from ≤ 10 individuals, making comparison difficult. Overall, it seems likely that amplitudes of circadian variation in vital signs are attenuated by some combination of pathology, treatment, and inactivity, with each vital sign responding differently.

Variability between demographic cohorts
As previously mentioned, Figs. 2 and 3 largely show the expected age-related increase in mean SBP and decrease in mean HR [34,44]. That these results are consistent across databases, and with results reported for non-ICU cohorts in the literature provides further support to the notion that the behaviour being observed is physiological behaviour, rather than behaviour governed by environmental influences.
However, mean SBP in men does not show age-related variations despite these being well documented in healthy men and present for women in the selected cohort [34]. It is important to note that the ICU population for a given demographic group is not necessarily representative of the general population for that demographic group, and this is elaborated further in Additional file 3. Broadly, young men (between 15 and 44 years) have a relatively high prevalence of admission diagnoses codes for HIV, alcohol abuse, and trauma not seen in younger or older women, or in older men. These variations in cause of ICU admission, and thus patient condition and treatment, may explain this lack of expected trends with age in mean SBP for men.

Variability between databases
As previously mentioned, PICRAM shows both an increased retention rate in the selected cohort (Table 3) and an elevated mean HR and SBP (Fig. 1). It is likely the increased retention rate of PICRAM ICU stays relative to MIMIC-III or eICU-CRD is due in part to the lack of discharge destination coding in the UK, leading to all patients expected to make a full recovery in PICRAM being retained, as opposed to only those discharged home as in MIMIC-III and eICU-CRD. Thus, it is likely that the increase in mean HR and SBP observed in PICRAM is due to the PICRAM cohort being older and more ill, rather than local variables or changes in clinical practice. These observations correspond to data present in the literature that suggest that patients in UK ICUs are on average more ill than those in US ICUs, associated with the lower number of ICU beds per capita available in the UK [45,46]. It is important to note that the circadian pattern shapes and intra-database trends with gender and age hold across all three databases, regardless of differences in clinical behaviour or shift timings, suggesting these profiles are widely generalisable.

Limitations
This study has several limitations that are worth discussing. All trends reported in this paper are for the average of large numbers of vital-sign measurements across a reasonably diverse cohort of patients. Thus, while the trends observed match trends reported for healthy individuals outside the ICU, and the trends are generally maintained when the data are broken up into subgroups by gender or age, there is little indication as to how consistently these trends can be observed on an individual basis. This is important as any prospective tracking of patient recovery, or of the development of complications such as delirium [5], would require the ability to meaningfully establish an individual's vital-sign circadian rhythms using routine clinical measurements.
While a large amount of patient data has been gathered across a large number of different hospitals, it also worth noting that data are still only gathered from 2 countries with lifestyles and demographics that are reasonably similar. Thus, further work is required to assess the generalisability of any trends observed to other countries where diet, clinical practice, and cause of ICU admission may vary to a greater extent.
This paper does not compare circadian rhythmicity between patients who 'recovered' and those who died. As such, the paper does not provide evidence of the 'sensitivity' of observable circadian vital-sign patterns to patient recovery, only that this behaviour can be observed in those who recovered. Research into generalisable circadian vital-sign behaviour in the ICU is relatively new. As such, it is important to establish that generalisable circadian behaviour exists prior to discharge in ICU patients who recovered, thus laying the groundwork for future comparisons between cohorts.
This paper does not demonstrate loss of circadian rhythmicity in the selected cohorts earlier in their ICU stay. Instead, it relies on existing literature that suggests circadian rhythms are severely disrupted in an ICU [3,4,15,23]. Patients early in an ICU stay are likely to have their vital-sign patterns directly disrupted by medication and clinical interventions, making observation of any underlying circadian pattern, whether present or not, significantly more challenging. Finally, this paper seeks to evaluate circadian rhythmicity in the last 24 h prior to discharge from an ICU in the typical ICU patient who recovered. As such, the relatively short stay of ICU patients in the US databases, attributable to broader intake criteria used in US ICUs [45,46], should be noted.

Conclusion
This paper investigated the presence of, and the relationships between, circadian rhythms in SBP, HR, RR, and T across a subset of patients in the MIMIC-III, eICU-CRD, and PICRAM ICU databases deemed most likely to exhibit circadian behaviour. Circadian patterns in SBP, HR, RR, and T that visually corresponded to those reported in the literature for non-ICU cohorts were observed, and these circadian patterns showed strong correlations between databases (mean R of 0.89). The peak-nadir excursions of the observed circadian patterns were reduced by a factor of 2-5 compared to behaviour reported in the literature for young, healthy individuals. These results support the existence of circadian rhythms in ICU patients who are within 24 h of discharge, and the generalisability of these circadian patterns across different cohorts subject to different standards of clinical practice. The existence of a generalisable circadian state prior to ICU discharge in patients who recovered has potential application in both prospective and retrospective tracking of patient recovery in the ICU, as well as the development of complications such as delirium.
Vital-sign circadian rhythms in patients prior to discharge from an ICU: A retrospective observational analysis of routinely recorded physiological data

Medication Exclusion Criteria
Overall, the goal of the medication exclusion criteria was to exclude medication that met the following criteria: • Would have a significant, rapid onset effect on patient vital signs, such that underlying trends in vital signs would be masked. • Would potentially be given to a significant proportion of patients within the last 24 hours prior to discharge from an ICU. Where possible, chronic, or longer term medication which was unlikely to significantly affect vital-sign trends hour-to-hour were not excluded. Additionally, administration of such chronic medication is often temporarily halted while patients are in an ICU and only restarted after discharge. Inclusion of patients on chronic medication was important in providing a reasonable picture of a recovering ICU patient, and retaining sufficient data for analysis. β-blockers, which are used orally as a treatment for chronic hypertension, were excluded when given intravenously, as this can have significant rapid onset effects on blood pressure and heart rate. The current exclusion criteria result in a reduction in vital-sign measurements of between 10.3% and 25.4% depending on the database (table 3 in the main text), thus the majority of patient data is retained. Figs. 1 -3 show the proportion of patients in the selected cohort who were treated by a broader list of vital sign altering medication within the last 26 hours (allowing for a~2 hour duration of action) of an ICU stay. It can be seen that the non-excluded medications administered most frequently to patients during this period were typically medications with longer-term effects (e.g. dexmedetomidinea sedative which, notably, does not cause respiratory effects, and diuretics, which act more indirectly on blood pressure). Fewer than 3% of patients in the selected cohort in each database are administered any given benzodiazepine or ACE inhibitor, and these medications are unlikely to dramatically alter vital signs in the short term. Calcium channel antagonists (e.g. Amlodipine, Diltiazem, Nicardepine) are more common, with 9.53% of PICRAM patients having Amlodipine and 8.68% of patients in eICU-CRD having Nicardipine in the final day.

Distribution of Measurements
This additional file shows the number of measurements averaged to create each of the 24-hourly data point plotted in the vital-sign profile figures. For all databases and vital signs, the number of measurements in each hourly bin that met the employed selection criteria was relatively consistent. MIMIC-III shows the clearest visible variations in frequency of measurements that met the selection criteria with time-of-day, with a maximum decrease in the number of such measurements by 32.0% -37.7% depending on vital sign (from peak) in the late afternoon. As the majority of patients are weaned off vital sign altering medication near to discharge from the ICU, there is a higher proportion of patients on such medication, for example, 20 -24 hours prior to discharge. As a large number of patients are discharged during the late afternoon, there is an increased likelihood of data being excluded from the late afternoon on the day before discharge. Interestingly, this reduction is not as clearly observable in the eICU-CRD data. Regardless, there is a significant and broadly similar amount of patient data available at any given time-of-day for each vital sign within each database.  Vital-sign circadian rhythms in patients prior to discharge from an ICU: A retrospective observational analysis of routinely recorded physiological data

Admission diagnoses
Figs. 1, 2 and 3 present the 15 most common diagnosis codes for different age and gender subgroups. Note that MIMIC-III and eICU-CRD used the ICD-9 coding system for admission diagnoses, while PICRAM used the ICNARC Coding Method (ICM). All analysis of diagnosis codes in this section is performed using up to the top 3 diagnosis codes for each patient. Note that the normalised frequency plotted is the proportion of ICU admissions, not admission diagnosis codes, that match a given admission diagnosis. Thus, for up to three diagnosis codes per patient, the total sum of normalised frequency will be between 1 and 3, rather than equal to 1.
From these figures, it is apparent that the younger 15 -44 year cohorts of both genders have notably more varied causes of admission than the older 45 -90+ year cohorts. Men in the younger cohort had the highest prevalence of diagnosis codes related to alcohol abuse, HIV, and trauma, by a considerable margin. Thus, younger men were one of the more heterogeneous cohorts, with a higher prevalence of HIV, alcohol, and trauma-related intake codes. Older men, in contrast, were a more homogeneous cohort. Women show a less dramatic variance in cohort heterogeneity with age, and younger women, while still heterogeneous, show a lower prevalence of alcohol-related, trauma-related, or otherwise unique admission diagnosis codes than younger men. This contrast may explain the lack of a clear upward trend in mean SBP with age in men, compared to the expected upward trend in mean SBP with age observed in women.