Cardiorespiratory dynamics measured from continuous ECG monitoring improves detection of deterioration in acute care patients: A retrospective cohort study

Background Charted vital signs and laboratory results represent intermittent samples of a patient’s dynamic physiologic state and have been used to calculate early warning scores to identify patients at risk of clinical deterioration. We hypothesized that the addition of cardiorespiratory dynamics measured from continuous electrocardiography (ECG) monitoring to intermittently sampled data improves the predictive validity of models trained to detect clinical deterioration prior to intensive care unit (ICU) transfer or unanticipated death. Methods and findings We analyzed 63 patient-years of ECG data from 8,105 acute care patient admissions at a tertiary care academic medical center. We developed models to predict deterioration resulting in ICU transfer or unanticipated death within the next 24 hours using either vital signs, laboratory results, or cardiorespiratory dynamics from continuous ECG monitoring and also evaluated models using all available data sources. We calculated the predictive validity (C-statistic), the net reclassification improvement, and the probability of achieving the difference in likelihood ratio χ2 for the additional degrees of freedom. The primary outcome occurred 755 times in 586 admissions (7%). We analyzed 395 clinical deteriorations with continuous ECG data in the 24 hours prior to an event. Using only continuous ECG measures resulted in a C-statistic of 0.65, similar to models using only laboratory results and vital signs (0.63 and 0.69 respectively). Addition of continuous ECG measures to models using conventional measurements improved the C-statistic by 0.01 and 0.07; a model integrating all data sources had a C-statistic of 0.73 with categorical net reclassification improvement of 0.09 for a change of 1 decile in risk. The difference in likelihood ratio χ2 between integrated models with and without cardiorespiratory dynamics was 2158 (p value: <0.001). Conclusions Cardiorespiratory dynamics from continuous ECG monitoring detect clinical deterioration in acute care patients and improve performance of conventional models that use only laboratory results and vital signs.


Introduction
When we admit acutely ill patients to the hospital, we make diagnoses and initiate therapies with an aim to restore or improve health. We do not intend for patients to deteriorate and thereby require transfer to an intensive care unit (ICU). Despite our intentions, escalation of care requiring transfer to an ICU occurs at a rate of 4 to 5 for every 100 acute care admissions. [1,2] Since mortality increases with every hour of delay in transferring critically ill patients to the ICU [3,4] the acute care community has focused on identifying patients at risk of potentially catastrophic conditions by developing and implementing Early Warning Scores (EWS) from data contained in electronic medical records (EMR). [1,2,[5][6][7][8][9] These EWS come in a variety of flavors, yet each comprises the same ingredients-intermittently measured vital signs (VS) and laboratory results (LABS) stored in the EMR.
A fundamental feature of current EWS is that they are updated only as new laboratory test results or vital sign measurements become available-this intermittent sampling may miss diagnostic patterns that would be revealed by sampling more often. Continuous physiological monitoring is often available as electrocardiographic (ECG) telemetry, but in current practice is only transiently displayed and then discarded. Analysis of continuous ECG reveals not only heart rates (HR) and dynamics such as variability, but can also extract respiratory rates (RR). [10,11] Thus, continuous ECG offers a window into cardiorespiratory dynamics, long known to hold information about the physiological status of the patient.
Here, we tested the hypothesis that the addition of cardiorespiratory dynamics measured from continuous ECG monitoring to conventional, intermittently sampled, data improves the predictive validity of models trained to detect clinical deterioration prior to ICU transfer or unanticipated death.

Study population
We studied consecutive adult (age > 18 years) patients admitted to acute care beds for whom continuous ECG data was available at the University of Virginia Health System, from October 11, 2013 to September 1, 2015. The 71 beds are arranged in 3 units and are under the care of a variety of hospital services, principally Cardiovascular Medicine and Cardiothoracic Surgery. An institutional electronic data warehouse archived the EMR including admission, discharge, and transfer information and a log of medical emergency team (MET) activations. ICU transfer, urgent surgery, death, or MET activation that preceded death. Two investigators reviewed unexpected deaths and transfers to the ICU for clinical deterioration. A third reviewer adjudicated cases in which there was disagreement.

Statistical analysis
We calculated the median values and interquartile ranges (IQR) for continuous variables and reported the counts and percentages of categorical measurements. During periods of continuous ECG monitoring, we developed multivariable prediction models to predict the risk of the primary composite outcome in the next 24 hours. We labeled observations during the time window prior to the event as cases, and observations outside this timeframe as controls.
In model development we adhered to the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD) statement recommendations. [23] To minimize the inclusion of redundant measurements and to avoid overfitting, we eliminated highly correlated predictors. We evaluated 48 candidate predictors derived from 39 component measurements (7 vital sign measurements, 25 lab results, 7 ECG calculations), including the slope or change of HR and ECG RR over windows of up to 24 hours. As new AF with rapid ventricular response commonly occurs and complicates critical illness, we also tested the interaction between current AF status and that of the initial rhythm detected. To impute missing values, we carried the last value forward for up to 24 hours for charted VS and for up to 48 hours for LABS; thereafter we imputed median values. We allowed our pre-specified candidate predictors to have non-monotonic and non-linear associations by modeling them with restricted cubic spline transformations. [24] We retained all significant higher-order factors. We adjusted for the repeated and correlated observations within individual patients with procedures that modify the variance-covariance matrix, and we quantified predictive accuracy using a concordance index (C-statistic). [24] We developed a first set of models from each independent data source (i.e. VS, LABS, ECG monitoring data). We conducted a second round of model development by modeling from all the pairwise combinations of independent data sources. Finally, we developed a combined model that integrated all available data sources. At each round to quantify the impact of adding the additional data source to the preceding model, we calculated the C-statistic and a categorical NRI for a change of 1 decile in risk. To test the hypothesis that the addition of cardiorespiratory dynamics to a model comprised of charted vital signs and laboratory results improves predictive validity, we calculated the probability of achieving the difference in the likelihood ratio χ 2 of the two models accounting for the difference in degrees of freedom between the two models and report this as a p-value. In sensitivity analysis we stratified the evaluation of model discrimination by the admitting hospital service. We also calculated several well-known EWS and compared their performance to that of our univariate and multivariate models.
We validated these models internally using bootstrap resampling (TRIPOD Type 1b model study) to estimate the performance on a new sample of observations from the same patient population, and report the bias-corrected C-statistics. [23] We justify bootstrapping methodology over split-sample or cross-validation techniques by noting that the estimates generated by these alternative strategies are less stable (different splits lead to different results) and exhibit greater bias (samples that vary only by chance will probably show similar performance). [23,25]

Characteristics of admissions
We studied 8,105 consecutive admissions with available continuous ECG data. After exclusion of observations occurring after DNR, DNI, or CMO orders, we analyzed the remaining 2.2 million observations (63.3 patient-years; Fig 1). Table 1 provides descriptive statistics of the cohort. The majority of patients were admitted to either Cardiovascular Medicine or Cardiothoracic Surgery services (5,504 admissions [68%], 81% of all patient-days). The median age was 65 (IQR: 55-75; range: 18-99) years and the median hospital length of stay was 4 (2-7) days. There was good agreement of HR and RR from the charted VS and the ECG analysis. RR from charted VS was most frequently one of three values: 18 (15.0%), 20 (11.7%), and 16 https://doi.org/10.1371/journal.pone.0181448.g001 (9.9%) breaths per minute (Fig 3). The distribution of ECG RR, as expected, was more continuous. AF was the first detected rhythm in 655 admissions (8.1%) and was present in 9.3% of all observations during ECG monitoring.

Outcomes
The primary composite outcome of deterioration leading to unanticipated ICU transfer or death occurred 755 times in 586 admissions (7%). Many events triggered MET activations (n: 341; 45.2%). While 159 patients died within 24 hours of last time recorded in a telemetered bed, only a minority were unanticipated and not accompanied by new DNR, DNI, or CMO orders (n: 30; 4% of all events). Admissions with events had 3-fold longer median hospital lengths of stay (11 vs 4 respectively) and a 43-fold increase in hospital mortality (17.2% vs 0.4% respectively; Table 1). Continuous ECG data was available prior to 395 of these events (52%) with which we developed new statistical models.

Models derived from individual data sources
A model derived from intermittently measured VS had a bias-corrected C-statistic of 0.687 (optimism: 0.008) and all predictors but temperature were associated with clinical deterioration (Fig 4). A model derived from intermittently measured LABS had a C-statistic of 0.629 (optimism: 0.025). A model derived from cardiorespiratory dynamics featuring HR, ECG-RR, AF, and the slopes of both HR and ECG-RR had a C-statistic of 0.651 (optimism: 0.010). cardiorespiratory dynamics and VS respectively (Fig 5). Adding cardiorespiratory dynamics to the LABS model resulted in a C-statistic of 0.704 (optimism: 0.011) with NRI of 0.14 and 0.16 for LABS and cardiorespiratory dynamics respectively. Combining VS with LABS resulted in a C-statistic of 0.714 (optimism: 0.014) with NRI of 0.20 and 0.13 for VS and LABS respectively. A fully integrated model using all available data sources had the best C-statistic: 0.729 (optimism: 0.014) with NRI of 0.15, 0.13, and 0.09 for the addition of VS, LABS and cardiorespiratory dynamics respectively. The difference in likelihood ratio χ 2 between integrated models with and without cardiorespiratory dynamics was 2158 with a difference of 13 degrees of freedom (p value: <0.001). The most important predictors were RR, supplemental oxygen use, SpO 2 , and ECG signal to noise ratio (ECG-SNR).

Effect of admitting hospital service
The incidence of clinical deterioration ranged from 1.15 to 2.24 per 100 monitored patientdays ( Table 2). While there were differences in model discrimination amongst various hospital services, as additional data sources were integrated into the model, and consistent with our more general analysis, the predictive validity incrementally improved. The fully integrated model had C-statistics of 0.715, 0.716, 0.775, and 0.796 for Cardiovascular Medicine, Cardiothoracic Surgery, Other Medicine, and Other Surgery respectively. The minimal anonymized data set necessary to replicate regression model analysis above is publicly available on the  Saturation or transparency depicts the statistical significance as calculated by the square root of (χ 2 minus degrees of freedom (df)) and can be likened to the absolute value of a Z score adjusted for the degrees of freedom associated with a predictor due to higher order terms. Hue or color represents the data source: individual data sources (green), pairwise combinations (blue), fully integrated model using all available data (red). Predictors that failed to achieve statistical significance in any of the models are not displayed.

Comparison to other early warning scores
In evaluating all univariate predictors, our newly derived models, and common EWS, we found that the top 5 performing models for the composite outcome were those newly developed and presented in this paper with the exception that the electronic Cardiac Arrest Risk Triage (eCART) score [8] outperformed our model derived from only continuous ECG metrics, and both the National Early Warning Score (NEWS) [7] and the VitalPac Early Warning Score (ViEWS) [6] outperformed our model derived only from LABS ( Table 3). Systemic Inflammatory Response Syndrome (SIRS) [26] came next as the 11 th best model evaluated while qSOFA [9] and Acute Physiology and Chronic Health Evaluation (APACHE) II [27] both performed more poorly than the univariate predictor of mean HR taken from either the continuous ECG monitor or from the EMR charted VS. All univariate predictors, models, and Early Warning Scores were evaluated in identical fashion where observations within 24-hours of the primary outcome were labeled as cases and all other at-risk observations were labeled as controls. Ranking is from best discrimination to worst as quantified by Cstatistic (non bias-corrected). We also report the degrees of freedom and the total statistical significance or In panel A, developing shock is recognized by models derived from multiple sources, but charted VS alone would have been sufficient to identify this patient several hours before ICU transfer. Panel B provides an example of a patient with developing confusion and acute renal failure where laboratory results but neither of the other data sources identified her trajectory. Panel C represents a patient on post-operative day #3 from an uncomplicated vascular surgery whose progressive abdominal pain and tachycardia prompted CT imaging and a lab draw, both of which were consistent with intra-abdominal hemorrhage leading to urgent exploratory laparotomy. ECG monitoring of cardiorespiratory dynamics identified this deterioration long before the other diagnostic tests were obtained that prompted urgent surgery. Panel D represents a patient with progressive respiratory failure and renal failure despite therapeutic drainage of a pleural effusion. While all data sources identified her as at increased risk, the integration of all data sources accentuated the severity and acuity of her deterioration.

Discussion
We studied continuous physiological monitoring prior to adverse events in patients hospitalized in acute care beds. We found that cardiorespiratory dynamics measured from continuous ECG monitoring improved the predictive validity of models based on intermittent measurements of VS and LABS. While multiple parameters helped to detect the deteriorating patient, RR, supplemental oxygen use, and SpO 2 were the most important. These findings underscore the need for accurate counts of RR and assessment of respiratory status. The quality of the ECG signal, measured as the ECG-SNR, was also important; underscoring that careful attention to subtle changes in skin properties-like temperature, dryness, and texture-may provide warning of impending patient decline. Detecting the deteriorating patient on the hospital ward is a major goal. The causes are varied, and range from underestimation of the admission diagnosis to the development of new and unrelated illnesses. The consequences range from altered management plans on the ward to ICU transfer and in some cases even to cardiac arrest and death. Physicians and nurses agree that early warning signs are often present, but are sometimes recognized only in retrospect. Beginning with APACHE for ICU patients in 1981 [28] and EWS for acute care patients in 1997 [29], many systems have been developed using charted elements-VS, level of consciousness, laboratory test results. In 2008, Smith and coworkers reviewed more than 60 such systems using single or multiple predictive parameters. [30,31] These systems benefit from systematically examining all the available data, and from the clinical wisdom of their developers. They lack, however, information from continuous physiological monitoring that is available for many hospitalized patients.
We hypothesized that the cardiorespiratory dynamics held information about the changing state of the hospital patient vulnerable to deterioration leading to ICU transfer, and we tested this hypothesis using advanced ECG signal processing techniques and a large, well-annotated database. From the ECG waveform signals, we determined linear and non-linear dynamical properties, identified AF, and detected respiration. From individual review of charts, we selected only patients whose ICU transfer was prompted by clinical deterioration, as opposed to elective transfer such as after cardiac surgery. The robustness of these approaches enhances, in our view, the validity of the results.
Our findings may be viewed in light of a paradigm shift in approaches to early detection of the deteriorating patient. This work began in intensive care units. In the neonatal ICU, advanced signal processing and HR time series analysis are used for early detection of neonatal sepsis, where it saves lives [32][33][34], as well as for other adverse outcomes. [35] In adult ICUs, signatures of illness have been identified in sepsis, respiratory failure leading to urgent intubation and hemorrhage leading to multi-unit transfusion of red blood cells. [36,37] In adult ICU patients already known to be septic, shock may be predicted. [38] The idea is spreading to the problem of early detection of the deteriorating floor patient. Zimlichman and coworkers have introduced contactless motion detection to derive HR and RR. [39][40][41] Tarassenko and coworkers developed a probability-based model of patient status from distributions of VS. [42] Hu and coworkers have used advanced data mining to analyze monitor alarms to produce SuperAlarms. [43] Churpek, Edelson and coworkers have devised EMR-based alerts of increasing risk of in-hospital cardiac arrest. [2] In this work, we introduce advanced time series analysis of continuous ECG telemetric monitoring to capture physiological dynamics in acute care patients. We captured these dynamics in the time, frequency, and non-linear dynamical domains. [14,15,17,18,44] We see many opportunities to apply these and other mathematical approaches to large, well-annotated clinical data sets to develop sophisticated and useful monitoring that use all available data.

Limitations
This is a single-center observational study in which Cardiovascular Medicine and Cardiothoracic Surgery were the predominant provider teams and thus our results may not generalize to other environments. MET or Rapid Response teams are varied in their composition and implementation across institutions. Because MET activation at our institution is not restricted to objective criteria, we included only MET activations that were followed by ICU transfer. We employed advanced logistic regression models, but note that ensemble methods in machine learning are well-suited for classification of heterogeneous groups.