Day-to-day progression of vital-sign circadian rhythms in the intensive care unit

This study employed the following three clinical databases:

Data from each database were used to build two cohorts of patients. The first cohort included patients who recovered from the ICU to be discharged home (cohort SRV). The second cohort included patients who either died before hospital discharge or were discharged into hospice care (cohort DCS). These two cohorts were selected following similar criteria to those used by [12]:

For each ICU stay, all measurements in each 1-h period were averaged for each vital sign. This process avoids weighting data towards ICU stays where patients were more ill, and thus likely to have more regular vital-sign measurements. Hourly values were recorded left aligned. (For example, the mean of measurements between 1:00 am and 1:59 am was recorded as occurring at 1:00 am.) Most vital signs were measured at least hourly, with the exception of temperature in MIMIC 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.

We decided to analyse vital-sign circadian rhythms for up to 6 days from admission for each patient, as we were interested in the trajectory of these rhythms from admission and the relationship between this trajectory and patient outcome. This decision provided additional days for analysis when patients began to be discharged or died, but ensured at least 40% of each cohort remained throughout (see Fig. S1 in Additional file 2), ensuring that a significant proportion of data remained available for analysis. We divided the 6 days of analysis into three periods, shown in Table 1. These periods are labelled in subsequent figures.

The association between these three periods and the number of available patient measurements can be seen in Fig. S2 in Additional file 2. Period III spans 48 h as a patient with a 4.0 day LOS admitted at 00:00 on the first day will be discharged or die at the beginning of day 4, 48 h before the end of the analysis period. Note that for patients with a LOS greater than 6.0 days, the monitoring period is the first 6.0 days of ICU stay.

Given the established variation in vital signs with age and gender [12, 23], the SRV and DCS cohorts were further broken into both gender and age subgroups. The age subgroups employed were based on those described for ‘health, health services and nutrition—morbidity and handicaps’ in ‘Provisional guidelines on standard international age classifications’ [24] and identical to those in Davidson et al. [12]: 15–44 years, 45–65 years and 65+ years of age.

In addition to basic demographic information such as median age or LOS, we also determined the Oxford Acute Severity of Illness Score (OASIS) [25]. OASIS is a severity of illness score used for predicting patient outcomes and readily determinable using retrospective patient databases as in this study. Common ICU severity of illness scores such as APACHE and SAPS requires a wide variety of information, such as past medical history and comorbidities, which are not necessarily recoverable from retrospective databases, and are thus less suited to this task.

Quantitative assessment of circadian rhythmicity was performed at the cohort level in two ways. First, peak–nadir excursions [26], which provide an indication of circadian rhythm amplitude, were computed for each cohort and each day as:where \(VS_{n}\) is the set of 24 h vital-sign values for the nth day. Second, the Pearson’s correlation coefficient (R) was computed between daily mean 24 h vital-sign profiles within a given cohort. The correlation was computed between each daily vital-sign profile and a demographically matched ‘recovered’ profile. This ‘recovered’ profile was created as follows for a given target cohort: Correlation coefficients were calculated as:where \(VS_{n}\) is the set of 24 h vital-sign values for the nth day and \(VS_{r}\) is the set of 24 hourly demographically matched ‘recovered’ vital-sign values. These quantitative metrics were not calculated for the first day after admission (period I) due to the variable number of measurements available during this period.

Figures 1, 2 and 3 show that vital-sign circadian rhythm profiles typical of non-ICU cohorts were present at the cohort level throughout an ICU stay in both patients who recovered and those who did not, though with a suppressed amplitude for the latter group. This result extends the findings of [12], which established the presence of typical vital-sign circadian rhythm profiles in the final day of ICU stay for patients who recovered. In Figs. 1, 2 and 3, consistent elevated SBP, HR, RR and temperature values during the day with peaks and nocturnal troughs were present. Typical vital-sign circadian profiles were not consistently observable in PICRAM, especially for SBP and the smaller DCS cohorts. This lack of observable vital-sign circadian profiles was likely due to the dearth of available data (see Table 3) rather than an underlying difference between the patient cohorts. However, it is worth noting that patients in PICRAM tended to have a greater OASIS score (Table 2) than those in MIMIC-III or eICU-CRD, reflecting the smaller number of ICU beds per capita available in the UK compared to the USA [27, 28]. The number of available measurements on the first day (period I, during which patients were admitted) varied significantly over the day (see Fig. S2 in Additional file 2).

The results shown in this paper are for grouped cohorts of patients, excluding periods of medication that significantly affected the vital signs analysed. 24–34% of patients in the DCS cohort died or were discharged to hospice care in the 4–6-day window (see Fig. S1 in Additional file 2). Thus, the majority of patients remained in the ICU throughout the reporting period. This means that the DCS cohort profiles presented are not representative of individual patients who are heavily medicated and within hours of dying. Such patients would be expected to exhibit significantly altered behaviour, including, potentially, a lack of observable vital-sign circadian rhythms.

There were clinically expected differences in mean levels between the SRV and DCS cohorts for most vital signs. The DCS cohorts consistently showed decreased mean SBP, increased mean HR, increased mean RR and decreased mean temperature, corresponding to the higher sensitivity periods of common early warning scores [29]. These trends were overlaid with expected gender (Fig. 1) and age related (Figs. 2, 3) trends in the mean vital signs, as shown in Davidson et al. [12]. Similar to the results in Pimental et al. [30], mean SBP tended to increase over the course of an ICU stay, and mean HR to decrease, with both of these trends being stronger in the SRV cohort. RR and temperature showed no consistent trends over the course of an ICU stay.

The vital-sign circadian rhythms in our analysis correspond to those reported in the literature for non-ICU cohorts in controlled environments [11, 13, 14], with a suppressed amplitude. They are present across different databases, age and gender cohorts. The three databases included in this study show data gathered over a long period of time at a single centre (MIMIC-III), data gathered over a shorter period of time at a variety of centres (eICU-CRD) and data gathered at a pair of centres in a different country with different standards of care (PICRAM), all of which show broadly comparable circadian patterns and inter-cohort trends. This suggests that the observed patterns are inherent circadian rhythms and not a product of the ICU environment. Cohorts where vital-sign profiles are difficult to interpret are consistently those where there is limited data available, rather than any consistent demographic or treatment-based factors. Thus, these results suggest that vital-sign circadian rhythms are present at a cohort level, throughout an ICU stay.

Figure 4 shows that, in general, peak–nadir excursions were greater for the SRV than the DCS cohort in MIMIC-III and eICU-CRD. The peak–nadir excursions in SBP, RR and HR tended to increase over the course of an ICU stay in the SRV cohort, while this behaviour was less prevalent in the DCS cohort. Peak–nadir excursions in vital-sign circadian rhythms have previously been shown to be suppressed in the ICU [12]. Thus, this observed increase in peak–nadir excursions in the SRV cohort over an ICU stay is potentially indicative of a ‘recovery’ of the magnitude of vital-sign circadian rhythms. While the average age of the SRV and DCS cohorts was significantly different (see Table 2), no significant differences in peak-nadir excursions between age groups were noted in Davidson et al. [12], so differing demographics on their own were unlikely to account for the greater peak–nadir excursions in the SRV cohort. Additionally, greater peak–nadir excursions in the SRV cohort compared to the DCS cohort were still present in the age subgroups (see Additional file 4).

It is important to note when interpreting peak–nadir excursions for grouped cohorts of patients that there are distinct known vital-sign patterns in, for example, SBP [6] that are associated with worsened long-term health outcomes. Thus, the lessened peak–nadir excursions in the DCS cohort may represent a cohort with a more diverse range of distinct rhythm shapes or frequencies, as opposed to a homogeneous cohort with lessened peak–nadir excursions. Future work could further cluster patients based on such vital-sign topographies, treating these groups as distinct.

The greater peak–nadir excursions in the SRV cohort compared to the DCS cohort in MIMIC-III and eICU-CRD are not observed for PICRAM. However, the greater peak–nadir excursions observed in the PICRAM DCS cohort, compared to the SRV cohort, are likely due to the relatively small number of measurements available resulting in noisy vital-sign profiles, rather than underlying differences in patient behaviour (see Table 3). This result highlights the weakness of peak–nadir excursions as a metric to quantify circadian rhythm strength in smaller cohorts of patients. Further evidence of the increase in peak–nadir excursions due to noise as cohort size is reduced is provided by the fact that eICU-CRD, the largest database, tended to have the clearest distinction between SRV and DCS cohort peak–nadirs. More evidence still is that the smaller age (15–44) and gender sub-cohorts in each database tended to have less clear distinction in peak–nadirs between the SRV and DCS cohorts (see Additional file 4).

Figure 5 shows that the correlation between 24-h vital-sign profiles and the corresponding recovered final day vital-sign profile provides a potentially more robust metric of circadian rhythmicity than peak–nadir excursions. The SRV cohorts had greater correlation in the majority of cases across all databases, with these correlations increasing over the course of the ICU stay. The DCS cohorts tended to have a lower correlation, and this correlation tended to further decrease after day 4, a trend most consistently visible in eICU-CRD (particularly for HR and RR). These trends were upheld with reasonable consistency across the age subgroups (see Additional file 4). However, the correlation metric penalises noise, as opposed to peak–nadir excursions which ‘reward’ it, thus the smaller size of the DCS cohorts lend themselves to a lower correlation.

Nevertheless, the differences in correlation and peak–nadir excursions between the SRV and DCS cohorts suggest a quantitative difference in circadian rhythms between the cohort of patients who recovered and those who did not. The attenuated amplitude and strength of these rhythms corresponds well to the results found in Davidson et al. [12], where in that study the survivors were the less well cohort and their peak–nadirs were shown to be attenuated relative to those of healthy individuals. This paper additionally shows the relationship between amplitude of circadian rhythms and well-being is monotonic: the strength of the rhythms for those who recovered is greater than for those who died. The combination of the results from the two analyses suggests that tracking vital-sign circadian rhythms throughout an ICU patient’s stay has the potential to provide additional prognostic information over the course of their hospital stay.

There are several limitations of note in this study. All the results presented are for cohorts of patients, rather than individuals. Any individual tracking of patient circadian rhythmicity would require significant additional development of the quantitative metrics presented in this paper.

The analysis presented in this paper necessarily excluded periods where the patient was under the effects of medication that might alter vital signs, such as vasoactive or inotropic medication. The exclusion of such periods was necessary as these treatments could potentially mask underlying vital-sign patterns. Further research would be required to account for the effects of these medications and thus construct patterns of circadian rhythmicity for patients during these periods.

There are demographic differences between the SRV and DCS cohorts. However, this is somewhat minimised by the use of gender and age specific sub-cohorts in the comparisons. Further, the tables in Additional file 3 show that medication and causes of admission are broadly comparable between the two cohorts.

The data employed in this paper are all from large retrospective clinical databases, as opposed to a study with a protocol for evaluating patient circadian rhythms. For example, this study considers the first 4–6 days of ICU stay, which allows us to investigate the initial trajectory of patient circadian rhythms in the ICU. However, patient LOSs varied considerably, with some of the included patients discharged between 4 and 6 days and others remaining in the ICU long after the 6 day mark. Additionally, this data is from two countries (the UK and USA) with somewhat similar demographics. However, this dataset still represents a large cohort spread over two countries and 211 hospitals, with a range of patients, standards of care and clinical practice.

In a retrospective study of a set of databases as large and diverse as these, there are a wide variety of possible approaches for creating and comparing cohorts. In this paper, the decision was made to choose the broadest comparison as a means of establishing the initial prognostic potential of vital-sign circadian rhythms over consecutive days of ICU stay. We compared two non-overlapping cohorts, one exhibiting a robustly coded ‘good’ outcome (the SRV cohort) and the other a robustly coded ‘poor’ outcome (the DCS cohort). Future work could explore a third cohort of patients who were discharged to long-term care facilities or rehabilitation centres. While some information on the association between patient LOS and circadian rhythms can be gleaned from the plots of consecutive days’ rhythms, and the longer average LOS of the DCS cohort compared to the SRV cohort, future work could more directly investigate this association.

Another potential area for future investigation is the complex relationship between sedation, wakefulness, circadian rhythms and delirium [31]. Vital-sign circadian rhythms exist separately from sleep and wakefulness, as shown by these rhythms being observed in studies where healthy volunteers were kept in state of sustained wakefulness with minimal activity [13, 14]. The inherently subjective nature of consciousness scores, combined with intermittent recording and the use of different scoring methodologies between different hospitals across multiple regions, makes it challenging to consistently establish and compare the wakefulness of patients [32]. Some level of standardisation in sleep patterns is inherently imposed by regular shift and meal times in the ICU. Similarly, the complex interaction between sedatives and consciousness makes establishing level of consciousness using medication administered difficult. Thus, categorising or classifying patients in this manner is significantly more difficult than categorising by outcome. Histograms of the most common medications administered to each cohort of patients, including sedatives, are shown in Additional file 3.

An important potential avenue for future work would be to look at the association between specific admission codes, disease states, treatments and circadian rhythms. For example, there are known associations between disrupted circadian rhythms and sedation [31] or conditions such as diabetes [33]. However, there is a significant variety of admission codes and treatments, and there are multiple challenges involved in selecting subgroups, overlap between these subgroups, and data sparsity. For example, the PICRAM database has a limited number of patients with the required instrumentation and length of stay that leads to very noisy vital sign profiles, without any sub-selection based on disease state. Histograms of the most common admission codes for each cohort of patients are shown in Additional file 3.

Ultimately, this study is able to conclude that, on average, patients who died had suppressed circadian rhythms relative to those who did not. The relatively narrow 95% confidence intervals of the means in the vital-sign profiles suggest that these population means are representative and significantly different between cohorts. However, we make no claims as to the cause of this observed effect. Thus, the conclusion of this paper is ‘circadian rhythms are suppressed in those who died relative to those who survived, and thus warrant further investigation as a potential metric of ICU patient’s condition’, but whether this suppression is due to certain aetiologies, ICU treatments, or environmental factors is not established.