HEART RATE MULTISCALE ENTROPY AT THREE HOURS PREDICTS HOSPITAL MORTALITY IN 3,154 TRAUMA PATIENTS

Abstract

Complexity is a measure of variation and randomness potentially indicating improvement or deterioration in critically ill patients. Previously, we have shown integer heart rate (HR) multiscale entropy (MSE), an indicator of complexity, predicts death based on long duration (12 h) and dense (≥0.4 Hz) windows of HR data. However, such restrictions reduce the use of MSE in the clinical setting. We hypothesized MSE predicts death using HR data of shorter duration and lower density. During the initial 24 h of intensive care unit stay, 3,154 patients had at least 3 h of continuous integer HR sampled. The first continuous window of 3, 6, 9, and 12 h was selected for each patient regardless of density, and an open-source MSE algorithm was applied (M. Costa, www.physionet.org; m = 2; r = 0.15). Risk of death based on MSE, alone and with covariates (age, sex, injury severity score), was assessed using randomly selected logistic regression in half of the cases. Area under the receiver operator curve (AUC) was computed in the other half in subgroups having various durations and densities of HR data. At days 2.3 (median) and 4.9 (mean), 441 patients (14%) died. Multiscale entropy stratified patients by mortality and was an independent predictor of death using 3 h or more of data. Multiscale entropy alone (AUC = 0.66 - 0.71) predicted death comparably to covariates alone (AUC = 0.72). We conclude: (1) Heart rate MSE within hours of admission predicts death occurring days later. (2) Multiscale entropy is robust to variation in bedside data duration and density occurring in a working intensive care unit. (3) Complexity may be a new clinical biomarker of outcome.

ABBREVIATIONS-AUC-area under receiver operator characteristic curve; HR-heart rate; HRV/HRV_i-heart rate/integer heart rate variability; ICU-intensive care unit; IRB-institutional review board; ISS-injury severity score; MSE-multiscale entropy; OR-odds ratio; SIMON-signal interpretation and monitoring; TRACS-Trauma Registry of the American College of Surgeons; VUMC-Vanderbilt University Medical Center

INTRODUCTION

Hemodynamic vital signs assist the clinician in determining onset and severity of shock, and end points of resuscitation, in the critically ill. However, more than a century after Cushing's (¹) ether charts, interpreting patient physiology remains largely a manual, intermittent, and subjective process. Human and technical limitations in processing vast amounts of clinical data have resulted in most second-by-second physiological measures displayed on bedside monitors being simply discarded.

This continuous physiological data, appropriately analyzed, may provide valuable information regarding the critically ill patient. For example, heart rate (HR) variability (²) has been studied in acute care settings (^{3, 4}) but requires continuous physiological data capture and analytic capabilities currently unavailable in most intensive care units (ICUs). We have studied integer HR variability (HRV_i) using a custom physiological data management system (⁵) and have previously shown that HRV_i is an early predictor of trauma patient mortality (⁶). In addition, HRV_i throughout the ICU stay is associated with death occurring several days later (⁷) and correlates with a number of traditional indicators of shock, including acidosis and coagulopathy (⁸), temperature extremes, adrenal insufficiency (⁹), and intracranial hypertension (¹⁰).

This study examines multiscale entropy (MSE) as a predictor of trauma patient mortality. Multiscale entropy measures variability in combination with the degree of randomness, or complexity, of HR (^{11, 12}) and may enhance the recognition of patients at risk for poor outcome. We have shown in a highly selected population of 285 patients meeting stringent data quality criteria that integer HR MSE predicts death as well as or better than HRV_i (¹³). We hypothesized that MSE would be a robust predictor of trauma patient mortality in the face of variations in data density and duration that may occur in a working ICU.

MATERIALS AND METHODS

Setting

This retrospective cohort study was performed at Vanderbilt University Medical Center (VUMC), the only level 1 trauma center serving a 65,000 square-mile area in the southeastern United States. Of the facility's approximately 3,500 annual trauma admissions, more than 1,800 are admitted to a 31-bed dedicated trauma unit. A subset of these beds comprise 14 trauma ICU beds accommodating 700 to 800 admissions per year. All trauma ICU beds are currently equipped with the signal interpretation and monitoring (SIMON) data capture system (⁵).

Patient care was delivered via VUMC's unique surgical critical care platform. The platform coordinates delivery of care to patients in VUMC's four surgical ICUs: trauma, burn, surgical critical care, and neurocare. Under the clinical component of the platform, all patient care is managed by a small number of faculty practicing under consistent evidence-based protocols. This system reduces variability in care and facilitates clinical research.

The research component of the platform comprises a staff providing 24/7 coverage who facilitate and simplify clinical research in the complicated surgical critical care environment. They are responsible for enrollment and consent of patients in studies, collection and management of specimens, and compliance with institutional review board (IRB) requirements. They bridge the clinical and research domains, helping to maintain a culture that supports research among the staff of the surgical ICUs.

Finally, the informatics component of the critical care platform captures and integrates data on patient physiology, pharmacology, clinical laboratory, demographics, and outcomes. When subjects are entered into a research study, these data are collected automatically, linked to specimen results, and merged into an IRB-approved deidentified repository for analysis. Overall, the critical care platform supports a wide variety of research projects. It reduces variability in practice and clinical outcomes, ensures standardization of data across numerous research projects, and fosters collaborations with nonclinicians who might otherwise have difficulty performing research in the critical care environment.

Data sources

Signal interpretation and monitoring

The SIMON system is an ongoing effort by the VUMC Division of Trauma and Surgical Critical Care to capture dense physiologic data from bedside medical devices. Signal interpretation and monitoring has captured trauma ICU patient physiologic data since December 2000, and captured parameters include HR, blood pressures, intracranial and cerebral perfusion pressures, arterial and venous oxygen saturations, core temperature, pulmonary and central venous pressures, cardiac index, and end-diastolic volume index. As of July 2007, data have been collected on more than 5,300 patients for their entire length of ICU stay in a SIMON-monitored bed, or more than 470,000 total hours of continuous monitoring and more than 7 billion data points. The SIMON data repository provided all HR data used in this study.

Trauma Registry of the American College of Surgeons

The VUMC Division of Trauma has participated in the Trauma Registry of the American College of Surgeons (TRACS) since its creation in 1996. All patients admitted to VUMC with trauma or burns are entered into this database, which includes all patients with SIMON data. Data are maintained locally and shared quarterly with the national repository. Currently, more than 300 parameters are captured into TRACS via retrospective chart review. In this study, outcomes, demographics, and injury severity scores (ISSs) were extracted from the TRACS data repository. After approval by the VUMC IRB, a deidentified data set was created linking information from the sources previously mentioned.

Study population

As shown in Figure 1, the study population is derived from 19,261 trauma admissions during a 63-month period who (1) were admitted to the trauma ICU (n = 4,456) and (2) had at least 3 h of continuous integer HR sampled during the initial 24 h of the ICU stay (n = 3,154). Continuous data were defined as the presence of more than 60 data points in each 5-min interval within the time period of interest. Furthermore, for inclusion, patients had to arrive in the ICU within 24 h of admission. Reasons for exclusion (n = 1,302) potentially include delayed arrival to the ICU, procedures or scans requiring trips off the unit or disconnection from monitoring equipment, early death or ICU discharge, or admission to a noninstrumented bed because only a subset of beds were instrumented through July 2003.

Fig. 1: Study population selection. Sufficient data were defined as 3 h or more of continuous HR in the first ICU day starting within 24 h of emergency department arrival. TICU indicates trauma ICU.

Measurements

Patient demographics (age in years, sex), acuity measurements (ISS, length of stay), and outcome (hospital mortality) were extracted from TRACS. The ISS is an anatomical scoring system that provides an overall score for patients with multiple injuries and has been shown to correlate with survival in trauma patients (¹⁴).

Integer HR data were automatically sampled by SIMON via RS-232 serial data export from bedside physiological monitors (Intellivue; Philips, Amsterdam, the Netherlands). The monitors report data every second. During the study period, SIMON dynamically adjusted sampling rate as a function of network load and other limits of then-current technology, resulting in integer HR data capture frequency between 0.25 and 1 Hz (data points recorded every 1 - 4 s). This provided an opportunity to investigate how variation in HR sampling frequency might affect the association between MSE and mortality. The monitor algorithm generating integer HR measurements is proprietary but likely represents a time-weighted average of several successive electrocardiogram interbeat intervals. For this study, patients were required to have at least 3 h of continuous integer HR sampled during the initial 24 h of ICU stay. The first available 3-h segment of data was selected for analysis. Additional HR data segments of 6, 9, and 12 h were extracted where such continuous data were available for a given patient.

Multiscale entropy was computed for each HR data segment using the publicly available algorithm of Costa et al. (¹¹) from PhysioNet (^{15, 16}) after interpolation of missing data. Using this method, sample entropy was calculated for a number of course-grained time series using a pattern length (m) of 2 and a similarity factor (r) of 0.15. As in our previous work investigating MSE's relationship to mortality (¹³), the sum of sample entropies was used as a summary measurement. Thus, each HR data segment (up to four segments for a given patient corresponding to the first continuous 3, 6, 9, and/or 12 h of continuous ICU HR data) has a corresponding single measurement representing MSE.

Statistical analysis

Demographics, acuity variables, body region injury scores, MSE, and mean HR were compared between survivors and nonsurvivors. Fisher exact test was used to compare categorical variables, and the Wilcoxon rank sum test was used for continuous variables. STATA (StataCorp, College Station, Tex) was used for all statistical computations, and a significance level of P < 0.05 was used throughout.

Two analyses were performed to assess the relationship of MSE and covariates to risk of hospital death at varying durations and densities of integer HR data. The first investigation was designed to test the independence of MSE measurements as predictors of mortality using data from as early as the first 3 h of ICU stay. A total of eight logistic regression models were created to assess the relationship between MSE and hospital mortality at each HR data duration (3, 6, 9, 12 h), with and without covariates (age, sex, ISS), that are established predictors of trauma patient outcome. To help guard against spurious results that might occur due to overfitting, a randomized test-validation protocol was used. Models were constructed using half of available cases selected at random, and predictive ability was evaluated by measuring the area under the receiver operator characteristic curve (AUC) obtained when fitting the model to the other half of cases. The odds ratio (OR) of MSE was computed to determine the corresponding change in risk of death for a unit increase in MSE (represented as the sum of SE over all scales).

To enhance the use in clinical settings, a second analysis was designed to study effects of varying data density and duration on predictive power of MSE. Cases were stratified into quartiles based on average sampling rate for each duration of HR data. Quartile break points were within 2% for all data durations, so the same breakpoints were used for convenience and to ensure comparability of results: 0.28 Hz (25th percentile), 0.40 Hz (median), 0.51 Hz (75th percentile). A single univariable logistic regression model for hospital death based on MSE was constructed using the longest, highest-frequency data group: 12 h of data sampled greater than 0.51 Hz. This model was then applied to all data groups, and the AUC was measured to determine goodness of fit. This analysis was repeated based on a model using data from the shortest, lowest-density data group (3 h of data sampled less than 0.28 Hz).

RESULTS

Demographic and HR characteristics are shown for survivors and nonsurvivors in Table 1. Mortality was 14%, amounting to 441 deaths. Those who died were significantly older and had more severe injuries (as measured by ISS) than survivors. Death occurred at a median of 2.3 days into hospital stay. Survivors' median length of stay was significantly longer at 8.2 days. Most patients were men, and whereas male sex seems to be associated with increased mortality, this effect was not statistically significant (P = 0.058) in univariate comparison. As in our previous studies, mean HR was not significantly different (P = 0.691), but MSE varied significantly between groups (P < 0.001).

Table 1: Population demographics and HR characteristics by survival

Survivors and nonsurvivors differed in magnitude of injury to specific body regions (Table 2), most notably in scores for the head and neck region, with more than half of nonsurvivors having a score of 5 (life-threatening injury). Smaller differences were seen in other region scores, with survivors having slightly higher scores for abdomen, extremity, and external regions. However, interpreting these differences requires caution because patients often sustain injuries to multiple body regions, and the coding process is somewhat subjective.

Table 2: Body region injury scores by survival

Figure 2 illustrates differences in sample entropy over multiple scales using the first 6 h of continuous HR data in ICU stay. Nonsurvivors show lower sample entropy at each scale factor and therefore lower MSE (sum of sample entropies). Survivors had higher sample entropy values, indicating more variability of HR and fewer consistent repeating patterns. Integer HR of survivors is more random compared with HR of nonsurvivors.

Fig. 2: Multiscale entropy using the first 6 h of continuous ICU HR data. Multiscale entropy calculates sample entropy (y axis) for different scales (x axis). At each scale N, the signal is averaged in groups of N data points.

This effect was consistent when MSE was computed across various durations of HR data. Table 3 summarizes logistic regression models for mortality based on MSE using various durations of HR data with and without covariates. Multiscale entropy alone was predictive of mortality using as little as 3 h of HR data (AUC = 0.69), with a relatively consistent OR of 0.88 to 0.96. Thus, an increase in MSE over the intraquartile range observed in the population (9 - 22 = 13) corresponds to roughly a 2- to 5-fold reduction in risk of death (0.88¹³ = 0.19 or approximately one fifth, 0.96¹³ = 0.59, or approximately one half).

Table 3: Mortality models based on MSE and covariates using various durations of HR data

Furthermore, MSE was independent of covariates in all multivariate models and improved predictive ability over models based on covariates alone. Although sex was not statistically significant in univariate comparison (Table 1), it was a significant but weak predictor of mortality in logistic regression models incorporating MSE, age, and ISS for the 3- and 6-h durations of HR data and was included in multivariate modeling. For comparison, models were also built based on covariates alone in the same groups previously mentioned, with AUCs ranging from 0.70 to 0.73.

Finally, effects of data density and duration were examined by evaluating MSE-based models for mortality in subgroups stratified by average sampling frequency and duration of HR data. Figure 3 shows AUCs obtained for a single univariable model evaluated over various subgroups defined by sampling frequency and duration. The model was developed using data from the longest-duration, highest-density subgroup. Results obtained when the model was developed in the shortest-duration, lowest-density subgroup were nearly identical. Although increased sampling frequency improves MSE's predictive ability, MSE is still indicative of outcome (albeit weakly; AUC ≥ 0.61) even when approximately three of every four HR data points are missing. When average sampling rate was greater than the median (0.40 Hz), AUCs obtained were consistently greater than 0.70.

Fig. 3: Multiscale entropy mortality model performance at various durations and densities of HR data. Area under the receiver operator characteristic curve (AUC) is reported for various durations and densities of integer HR data. Density categories correspond to quartiles based on average sampling rate. Standard errors (not shown) ranged from 0.03 to 0.04 except for 12 h of medium- to high-density data, where the standard error was 0.06.

DISCUSSION

Patterns within continuously captured physiological data may signal unexpected deterioration or improvement in critically ill patients. Beat-by-beat analysis of HR has been used to assess autonomic function (¹⁷) and identify onset of sepsis (¹⁸), multiple organ dysfunction syndrome (¹⁹), and central nervous system injury (²⁰), to name a few recent examples. These techniques and other novel analyses applied to HR and other physiological signals may reflect system complexity, providing information that is independent from traditional variability metrics (²¹). Consequently, appealing hypotheses regarding decomplexification (^{9, 22, 23}) and organ system uncoupling (^24-26) have emerged. Biological systems exhibit complex structure and function, and complexity theory provides a new paradigm for understanding, identifying, and treating shock (^{27, 28}) and critical illness (²⁹).

Strengths and limitations

This study builds on past efforts by our group and others to define the value of continuous physiological data, and concepts of biological complexity, in critical illness. It demonstrates, in the largest ICU study of adult HR dynamics to date, that continuous physiological data within hours of admission stratifies risk of death occurring days later. The AUCs obtained using only 3 h of data are roughly equivalent to those obtained at 12 h in our previous work, using simple measurements of HRV (⁶). Furthermore, the techniques described herein can be easily translated to the bedside. Integer HR data are efficient to capture and store, and MSE seems robust to variation in data duration and density that may occur in a working ICU. These results support the hypothesis that decomplexification signals critically ill patient risk and suggest a systems-analytic approach toward understanding pathophysiology of shock and critical illness.

Our infrastructure allows us to explore these concepts in day-to-day patient care. For example, each morning, attending physicians receive a report listing every trauma ICU patient's vital sign statistics, including HR dynamics, during the previous 24 h. Anecdotal observations suggest these measurements provide useful summary-level indicators of patient acuity. As we address limitations listed in the succeeding paragraphs and similar studies, we anticipate more formal evaluation and use of bedside decision support tools arising from measurements of physiological complexity.

Limitations of this study stem from its retrospective, observational nature and omission of potentially relevant clinical covariates. Although care of trauma ICU patients is generally similar early in the hospital course, therapy can differ based on severity of illness. If specific therapies (potentially including medications, procedures, mechanical ventilation settings, etc.) early in the hospital stay are indeed correlated with risk of death and also reduce MSE, they might contribute to the observed effects. In addition, incorporating information about cause of death and/or mechanism of injury may suggest underlying processes related to loss of complexity. Although this study illustrates that patients who died have more severe head and neck injuries, more work is needed to establish definitive links between mechanisms of injury, cause of death, and changes in complex physioregulation. As our clinical and research-based information systems become more integrated, we will be able to assess effects of additional clinical variables, which may strengthen or attenuate the relationships observed here and/or provide insight into underlying mechanisms of HR dynamics in the critically ill.

Finally, the use of integer HR data allows for inexpensive, automated data capture over large numbers of patients but limits the precision of HR measurement. Because these data represent time-averaged R-R intervals, sample entropy computation may be affected especially at smaller-scale factors when interpolation is performed. Relative changes in sample entropy between scale factors within individual patients may provide additional insight but likely requires more precise data available via electrocardiogram waveform analysis.

Future

This study explores a single new method of detecting clinically important patterns within a single densely sampled physiological signal. However, entropy is but one of many possible analytic methods reflecting complexity, and HR one of many possible signals, that may prove useful in assessing critically ill patient status. The analysis of high-density physiological data, not only cardiovascular parameters but also continuously monitored serum and genetic biomarkers, will become commonplace. Patterns within these vast data streams promise to illuminate specific disruptions in cells, organs, and systems. The resulting information, appropriately managed, will provide early warning of specific complications, hasten therapy, and inform a new generation of medical monitoring and decision-support strategies.

In the short term, we will refine the precision of our HR measurements and evaluate the ability of MSE and similar measurements to provide meaningful information using other physiological signals. Once validated, MSE and similar metrics must be integrated into medical decision-support processes. This transition requires new methods of capturing, linking, and delivering information to clinicians, an information environment we call the virtual bedside. We have begun to realize the virtual bedside though capture and analysis of dense physiological data. However, increasingly sophisticated tools such as MSE require substantial investment in informatics expertise and infrastructure to validate and deliver to clinicians appropriately, in real time. In the long term, such investment must be justified by multicenter, prospective, randomized, controlled experiments such as the HeRO trail investigating the impact of monitoring neonatal HR dynamics (³⁰).

Complexity measurements and other emerging algorithms combined with new clinical biomarkers promise to stratify high-risk patients, determine patient trajectory (improvement or deterioration), and uncover new disease mechanisms. These indicators must be studied and refined and then delivered to clinicians at the right place, time, and in the right format. Finally, such initiatives must be assessed to determine the degree they result in significant, general, cost-effective improvements to clinical decision making and quality of patient care.

CONCLUSIONS

(1) Heart rate MSE within hours of admission predicts death occurring days later. (2) Multiscale entropy is robust to variation in bedside data duration and density occurring in a working ICU. (3) Complexity may be a new physiological biomarker of outcome.

ACKNOWLEDGMENTS

The authors thank three anonymous reviewers for the constructive critique of earlier drafts. The physionet project (www.physionet.org) is gratefully acknowledged for providing the source code supporting this and other research efforts in complex physiologic signals.