Serial Heart Rate Variability Measures for Risk Prediction of Septic Patients in the Emergency Department

Abstract

In this study, we used serial heart rate variability (HRV) measures over 2 hours to improve the prediction of 30-day in-hospital mortality among septic patients in the emergency department (ED). We presented a generalizable methodology for processing and analysing HRV time series (HRVTS) data which may be noisy and incomplete. Feature sets were created from the HRVTS data of 162 patients with suspected sepsis using aggregation-based, deltabased and regression-based series-to-point transformations, and modelled over 100 random stratified splits. An optimized feature set comprising 12 selected HRVTS features performed better than baseline feature sets which only included patient demographics, vital signs and single time-point HRV measures taken at triage. This improved risk stratification approach could be used in the ED to identify high-risk septic patients for appropriate management and disposition.

Introduction

Sepsis is an increasingly recognised global problem that has a 10-20% in-hospital mortality (IHM) rate^1-3. Risk stratification of septic patients in the Emergency Department (ED) may help to guide appropriate management and disposition, thereby reducing morbidity and mortality^4-6. A number of clinical tools, such as the Quick Sequential Organ Failure Assessment (qSOFA), National Early Warning Score (NEWS) and Modified Early Warning Score (MEWS), have been used to risk stratify septic patients in the ED, where certain clinical information, such as laboratory investigations, are initially not available. These risk calculators rely solely on patient demographics, vital signs and clinical observations in their scoring criteria.

Several studies have reported the prognostic value of heart rate variability (HRV) parameters in septic patients presenting to the ED^7-9. Septic patients have reduced sympatho-vagal balance and impaired sympathetic activity, which lead to varying degrees of cardiac autonomic dysfunction¹⁰. This can be detected by HRV analysis, a quick, non-invasive technique of evaluating the beat-to-beat variation in heart rate. HRV analyses are divided into linear and non-linear methods¹¹. Linear methods include HRV parameters measured in time or frequency domains. Time domain HRV parameters are statistical calculations of consecutive R-R time intervals and how they correlate with each other. Frequency domain HRV parameters are based on spectral analysis. Studies have suggested that regulators of the cardiovascular system interact in a non-linear way^{12, 13} and HRV analysis using non-linear methods reflect these mechanisms¹⁴.

We previously demonstrated that HRV parameters measured at triage, combined with patient demographics and vital signs, can improve the prediction of 30-day IHM among septic patients in the ED compared to traditional risk stratification tools^15,16. Since sicker patients are often put on continuous electrocardiogram (ECG) monitoring in the ED, their serial HRV measurements over time can also be derived. We postulated that HRV time series (HRVTS) may embed more prognostic information than HRV parameters derived at a single time-point, such as at triage.

To date, there has been no published study exploring the use of HRVTS for risk prediction in sepsis. Thus, in this study, we aimed to compare the performance of HRVTS over 2 hours against single time-point HRV measures at triage, in the prediction of 30-day IHM among adult septic patients in the ED setting. In addition, since serial HRV measures are seldom encountered in literature, we aimed to develop and present a generalizable methodology for processing and analysing such data which may be noisy and incomplete.

Methods

Data preparation

Ethics approval for the study was obtained from SingHealth’s Centralised Institutional Review Board (CIRB, Reference Number 2016/2858), with waiver of patient consent. Patients above 21 years old who presented to the Singapore General Hospital (SGH) ED between September 2014 and April 2016 with clinically suspected sepsis, and who met at least 2 of the 4 Systemic Inflammatory Response Syndrome (SIRS) criteria¹⁷ were included. The SIRS criteria are temperature (<36°C or >38°C), heart rate (>90 beats/min), respiratory rate (>20 breaths/min) and total white count (<4000/mm³ or >12000/mm³).

Patient demographics (age, gender, ethnicity), first set of vital signs recorded at triage (temperature, respiratory rate, heart rate, systolic and diastolic blood pressures, and Glasgow Coma Scale (GCS) score), and outcome of interest (IHM within 30 days of ED admission) were retrieved from the hospital’s electronic medical records. Continuous one-lead ECG tracings over 2 hours in the ED were also obtained from X-Series Monitor (ZOLL Medical Corporation, Chelmsford, MA). The ECGs were split into 8 intervals of 15 minutes each and then analysed for HRV parameters using the most stable continuous 5-minute segment within each interval. 22 HRV measures in time, frequency and non-linear domains were computed for each interval. For patients whose ECGs were shorter than 2 hours or unanalysable due to non-sinus rhythm, excessive artifacts and/or ectopic beats, their data was considered missing for the relevant intervals.

To ensure data quality and to limit the amount of missing data in subsequent analysis, we only included patients with at least 7 of the 8 intervals in our final analysis (i.e. at most 1 missing interval allowed). Univariate statistical analysis was conducted to compare patients with and without outcome, as well as to compare the original and final cohorts (Kruskal-Wallis H test for continuous variables and Chi-squared test for categorial variables).

Feature Extraction

We created 2 baseline feature sets that did not utilize any information from the HRVTS as baselines for comparison. We then designed 3 groups of candidate feature sets (aggregation-based, delta-based and regression-based) using different methods of series-to-point transformations to encapsulate the HRVTS information. Detailed definitions of the feature sets, as well as an example of how the feature sets were calculated, are presented in Tables 1 and 2 respectively.

Table 1.

Definitions of the feature sets with feature values calculated based on the example mean_rr HRVTS.

Category	Feature set	Definition of features (additional to fs_no_HRV)	Name of HRV features	Feature value
Baseline	fs_no_HRV	Only gender, race, age and the vital signs	-	-
	fs_baseline	First 5-min HRV parameter at triage	mean_rr	510.00
Aggregation based	fs_first_last	First and last value of HRVTS	first_mean_rr	514.00
			last_mean_rr	488.00
	fsmedlQR	Median and inter-quantile range of HRVTS	median_mean_rr	501.00
			IQR_mean_rr	65.00
	fs_min_max	Minimum and maximum of HRVTS	min_mean_rr	406.00
			max_mean_rr	632.00
Delta based	fs_delta_first_last	First and last of all the differences in absolute value between neighbouring intervals	delta_first_mean_rr	58.00
			delta_last_mean_rr	46.00
	fsdeltamedlQR	Median and inter-quantile range of all the differences in absolute value between neighbouring intervals	delta_median_mean_rr	78.00
			delta_IQR_mean_rr	57.25
	fs_delta_min_max	Minimum and maximum of all the differences in absolute value between neighbouring intervals	delta_min_mean_rr	50.00
			delta_max_mean_rr	113.00
Regression based	fs_reg_default	Coefficient and intercept of the linear regression model fit on the HRVTS	coef_mean_rr	10.86
			intercept_mean_rr	463.57
	fs_reg_delta	Coefficient and intercept of the linear regression model fit on the series of the differences in absolute value between neighbouring intervals	coef_delta_mean_rr	2.40
			intercept_delta_mean_rr	71.27
	fs_reg_cum_delta	Coefficient and intercept of the linear regression model fit on the series of cumulative sum of all the differences in absolute value between neighbouring intervals	coef_cum_delta_mean_rr	91.00
			intercept_cum_delta_mean_rr	-46.67

Table 2.

Example of a patient’s mean_rr HRVTS and derived values for the calculation of features. The 4th interval was missing, and the interpolated value is indicated by an underline.

Time	Mean_rr	Derived from raw value		Absolute delta	Derived from absolute delta		Cumulative	Derived from cumulative absolute delta
1	514.00	median	501.00	58.00	median	78.00	58.00	Linear Regression	y = 91.00x-46.67
2	456.00			50.00			108.00
3	406.00	IOR	65.00	113.00	IOR	57.25	221.00
4	missing			113.00			334.00
5	623.00	Linear Regression	y = 10.86x+463.57	98.00	Linear Regression	y = 2.40x+71.27	432.00
6	534.00			46.00			478.00
7	488.00			-			-
Baseline	510.0			-			-

(HRVTS, heart rate variability time series; mean_rr, average width of the RR interval)

Baseline feature sets: The first baseline feature set (fs_no_HRV) contained only patient demographics and vital signs. These features are included in all subsequent feature sets. In the second baseline feature set (fs_baseline), single time-point HRV parameters generated from the first 5-minutes of ECG obtained at triage were added.

Aggregation-based features sets: 3 aggregation-based feature sets were constructed by taking (1) the first and last values of the HRVTS (fs_first_last), (2) the median and interquartile range (IQR) of the HRVTS (fs_med_IQR), and (3) the minimum and maximum values of the HRVTS (fs_min_max).

Delta-based feature sets: With the assumption that the magnitude of relative changes between neighbouring intervals were informative, we calculated (1) the first and last (fs_delta_first_last), (2) the median and IQR (fs_delta_med_IQR), and (3) the minimum and maximum (fs_delta_min_max) of the absolute differences between adjacent HRVTS measurements. Linear interpolation was used to impute missing measurements.

Regression-based feature sets: With the assumption that trends in the HRVTS were informative, we fitted the (1) original values (fs_reg_default), (2) absolute differences between measurements (fs_reg_delta), and (3) cumulative differences between measurements (fs_reg_cum_delta) of the HRVTS to a linear regression model using the ordinary least squares method. The linear coefficients and intercepts obtained were used as features to characterize the trends.

Model Setup

Prior to modelling, all features were standardized to be between 0 to 1 by MinMaxScaler (subtracting the minimum value and then dividing by the difference between the maximum and minimum values).

We set aside 20% of the patients as test data, stratified by outcome. Within the remaining 80% of the patients, we created 100 random stratified splits, with 80% as training data and 20% as validation data in each split. We chose logistic regression with L1 regularization (LASSO) with C=1.0 as our model due to its simplicity and the interpretability of odds ratios obtained. Class imbalance was addressed by applying sample weights inversely proportional to the number of samples. We used Area Under the Receiver Operating Characteristic curve (AUC) as our performance evaluation metric.

Feature Selection

Each feature set was inputted separately into the logistic regression model and the median odds ratios of all non-zero coefficients obtained across 100 splits were calculated. We tested the proportion of each feature having non-zero coefficient larger than 0.5 using one-sided binomial test under Bonferroni correction.

To create an optimized feature set, the feature with the most extreme odds ratio among statistically significant ones (if any) was selected for each HRV parameter. These selected HRV features were then combined with demographic and vital sign variables significant across majority of feature sets to form an optimized feature set fs_final_selection.

HRV analysis was performed using Kubios HRV software version 2.2 (Kuopio, Finland)¹⁸. Statistical analysis and modelling were carried out in Python 3.6 (Python Software Foundation, Wilmington, Delaware, USA) using the statsmodel¹⁹ and scikit-learn libraries²⁰.

Results

Figure 1 shows the cohort selection process. 214 patients were in the original cohort whose ECGs were split into intervals and analysed for HRV parameters. After excluding patients with more than 1 missing interval, 162 patients were included in the final cohort, of whom 31 (19.1%) met the outcome. There were no significant differences in the patient demographics, vital signs and outcome between the original and final cohorts (Table 3).

Figure 1.

Cohort selection flowchart.

Table 3.

Patient demographics and vital signs of the original and final cohorts used in the analysis.

		Original Cohort			Final Cohort			Original vs. Final
Characteristics		No 30-day IHM	30-day IHM	P-Value	No 30-day IHM	30-day IHM	P-Value	P-Value
n (%)		174 (81.3)	40 (18.7)	-	131 (80.8)	31 (19.2)	-	0.981
Age (years)		66.0 [56.2,76.8]	76.0 [67.8,83.0]	<0.001	66.0 [57.0,77.0]	76.0 [68.0,83.0]	0.002	0.806
Male gender		88 (50.6)	20 (50.0)	0.913	62 (47.3)	15 (48.4)	0.925	0.646
Ethnicity (%)
	Chinese	125 (71.8)	30 (75.0)	Ref.	97 (74.0)	25 (80.6)	Ref.	Ref.
	Malay	25 (14.4)	4 (10.0)	0.637	16(12.2)	4(12.9)	1.000	0.850
	Indian	15 (8.6)	5(12.5)	0.545	11 (8.4)	1(3.2)	0.465	0.631
	Others	9(5.2)	1(2.5)	0.692	7(5.3)	1(3.2)	1.000	0.901
Temperature (°C)		38.3 [37.4,38.8]	37.4 [36.5,38.5]	0.005	38.3 [37.2,38.9]	37.1 [36.0,38.7]	0.010	0.844
Respiratory rate (breaths/min)		19.0 [18.0,22.0]	22.0 [18.8,26.0]	<0.001	19.0 [18.0,21.5]	22.0 [18.5,24.5]	0.009	0.794
Heart rate (beats/min)		118.0 [105.0,130.0]	108.5 [97.8,125.5]	0.078	118.0 [105.0,131.0]	106.0 [97.0,113.0]	0.009	0.751
Systolic BP (mmHg)		114.0 [91.0,142.8]	105.0 [91.0,126.2]	0.097	115.0 [90.5,145.0]	102.0 [86.0,119.5]	0.042	0.926
Diastolic BP (mmHg)		63.0 [55.0,74.0]	61.5 [52.0,70.2]	0.442	63.0 [55.0,75.0]	61.0 [50.0,68.5]	0.204	0.995
GCS score		15.0 [13.2,15.0]	14.0 [9.0,15.0]	0.002	15.0 [14.0,15.0]	14.0 [9.0,15.0]	0.006	0.779

(BP, blood pressure; GCS, Glasgow Coma Scale)

For continuous variables, data is presented in medians and interquartile ranges. Kruskal-Wallis H test was used to test for differences. For categorical variables, data is presented in frequencies and percentages. Chi-square test was used to test for association.

Among the final cohort, 77 (47.5%) of the patients were male, with median age of 67 years (inter-quartile range, IQR 57–79). Patients who met the outcome were older (median age 76 years; IQR 68–83 years) than those who did not (median age 66 years; IQR 57–77 years). There were no significant differences in gender and ethnicity distributions between the two groups. In terms of vital signs, patients who met the outcome had higher respiratory rates, as well as lower temperatures, heart rates, systolic blood pressures and GCS scores, compared to patients who did not meet the outcome. The HRVTS of the two groups are visualised in Figure 2.

Figure 2.

Visualization of HRVTS for patients with outcome (orange) and without outcome (blue). (a) Time domain parameters (b) Non-linear domain parameters (c) Frequency domain parameters.

Table 4 shows the median odds ratios across all feature sets and predictor variables. Most of the features had coefficients of zero (hence odds ratio of zero) due to the L1 regularization. 12 HRV parameters (5 in time domain, 4 in frequency domain and 3 in non-linear domain) had statistically significant odds ratios and were selected into the optimized feature set fs_final_selection (Table 5). 3 other non-HRV variables (age, systolic blood pressure and GCS) were also statistically significant across most feature sets and added to the optimized feature set fs_final_selection to make a total of 15 features.

Table 4.

Median odds ratio of all the features. Green cells indicate odds ratio >1, and red cells indicate odds ratio <1. The darker the color, the stronger the odds ratio deviates from 1. All significant features are in black text. Values with a blue box are selected for fs_final_selection.

Mean RR, average width of the RR interval; SD RR, standard deviation of all RR intervals; HR, heart rate; RMSSD, root mean square of differences between adjacent RR intervals; NN50, number of consecutive RR intervals differing by more than 50 ms; pNN50, percentage of consecutive RR intervals differing by more than 50 ms; TINN, baseline width of a triangle fit into the RR interval histogram using a least squares; VLF, very low frequency; LF, low frequency; HF, high frequency; norm, normalized; LF/HF, ratio of LF power to HF power; DFA, detrendedfluctuation analysis; BP, blood pressure; GCS, Glasgow Coma Scale.

Table 5.

Features in the optimized feature set fs_final_selection

	Featurization	From Feature Set
Time Domain
Mean RR (s)	Minimum	fs min max
Mean HR (bpm)	Intercept	fs reg default
NN50 (count)	Last	fs delta first last
RR triangular index	Median	fsdeltamedIQR
TINN	Last	fs first last
Frequency Domain
VLF power (ms2)	Minimum	fs delta min max
LF power norm (n.u.)	Intercept	fs reg delta
HF power norm (n.u.)	Minimum	fs min max
LF/HF	Intercept	fs reg delta
Non-linear Domain
Approximate entropy	Coef	fs reg cum delta
Sample entropy	IQR	fs delta med IQR
DFA, a-1	Intercept	fs reg delta
Demographics
Age	Original	all
Systolic BP (mmHg)	Original	all
GCS score	Original	all

Figure 3 summarizes the median AUC of all feature sets over 100 training splits. The optimized HRVTS feature set fs_final_selection achieved the highest median AUC of 0.82, followed by the candidate feature set fs_min_max which achieved median AUC of 0.77. Both performed better than baseline feature sets which do not contain HRVTS information, namely fs_no_HRV (median AUC 0.68) and fs_baseline (median AUC 0.70). Figure 4 compares the median Receiver Operating Characteristic (ROC) curves of these four feature sets over 100 training splits. Table 6 reports the AUC achieved on the test set by these four feature sets. fs_final_selection outperformed the rest by a significant margin, while fs_min_max did not perform well, suggesting potential overfitting with single featurization method.

Figure 3.

Median AUCs of all feature sets over 100 training splits.

Figure 4.

Receiver Operating Characteristics (ROC) curves of fs_final_selection, fs_min_max and two baseline feature sets. AUCs reported in median ± IQR over 100 training splits.

Table 6.

Summary of median AUC achieved on validation sets and AUC achieved on test set.

Feature Set	Median AUC (Validation)	AUC (Test)
fs_no_HRV	0.68	0.57
fs_baseline	0.70	0.67
fs min max	0.77	0.63
fs_final_selection	0.82	0.76

Discussion

In this study, we used HRVTS data collected over 2 hours to improve the 30-day IHM prediction of septic patients on continuous ECG monitoring in the ED. An optimized feature set comprising 12 selected HRVTS features performed better than baseline feature sets which only included patient demographics, vital signs and single time-point HRV measures obtained at triage. This improved risk stratification could be used to inform management and disposition, such as early antibiotic therapy and ICU admission for identified high-risk patients.

While more research is needed to understand the physiological meaning and significance of each of the HRV parameters, studies have shown that HRV measures correlate with autonomic function²¹ and are predictive of adverse outcomes in septic patients. We hypothesize that higher-order dynamics and interactions in a HRVTS, including relative changes and trends over time may carry more prognostic information than the actual values of the HRV parameters themselves. This may explain why delta-based and regression-based features derived from HRVTS appear to be important risk predictors in our model.

We comprehensively explored various methods of feature engineering and selection to synthesize and make sense of the HRVTS data. In particular, our featurization method transforms a large number of data points for each patient into 15 key predictors and handles missing data. With the widespread adoption of wearable devices connected to the Internet of Things, we believe this type of high-resolution, high-volume time series data will become increasingly available. Therefore, the importance of our work lies in developing a methodology for processing and analysing HRVTS data which could be generalizable to other clinical settings and outcomes of interest, rather than in presenting a predictive model to be applied globally.

Our outcome of interest was IHM within 30 days during the same admission where the vitals and ECG were taken. Some studies did not specify a time period for mortality²² or if it was strictly within the same admission or not²³. We chose this endpoint as it is more likely to be sepsis-related compared to an out-of-hospital mortality or mortality from a subsequent admission. It is also more meaningful for physicians in terms of administering possible interventions such as closer monitoring and more aggressive management of high-risk patients²⁴.

Our study had several limitations. Firstly, this was a single-institution study with a small sample size. Larger multi-centre prospective studies are required to validate our results. Secondly, we had included patients in our study based on clinical suspicion of sepsis and meeting at least 2 of the 4 SIRS criteria. Sepsis largely remains a clinical diagnosis and there is no gold standard to determine whether a patient is septic. Other studies have attempted to address this issue by including only patients with administered intravenous antibiotics, blood culture investigations or confirmed source of infection^22,24. We acknowledge that our cohort definition reflects suspected sepsis rather than confirmed sepsis. However, we believe this is reasonable given our eventual application in the ED where laboratory test results and confirmed diagnoses may not yet be available. In addition, while the SIRS criteria has recently been replaced with a new state of sepsis, defined as a life-threatening organ dysfunction caused by a dysregulated host response to infection, the usefulness of the SIRS criteria in diagnosis of sepsis was still emphasized by the same task force25, ^26.

Thirdly, even though HRVTS can improve mortality prediction in suspected sepsis patients as shown in this study, they require more time and effort to obtain than single time-point HRV measures. Currently, we are developing a wearable device which can continuously monitor a patient’s ECG and automatically perform HRV analysis in real-time.

Fourthly, the time frame of 2 hours was chosen in this study based on the duration most patients spend in the ED while awaiting admission. Nevertheless, in our future studies, it is definitely worthwhile to explore the utility of HRVTS over a much longer period (such as 24-72 hours) and extending into the inpatient ward setting, as well as assess whether HRV-based risk predictions impact clinical management and outcomes. As continuous ECG data is inherently noisy (for example, due to patient movement or temporary disconnection), we chose to split the ECGs into 8 intervals of 15 minutes so that there is sufficient temporal window within each interval to capture a valid 5-minute segment for HRV analysis. However, this could have resulted in loss of information. Further study is needed to understand how we can better achieve granularity with noisy data.

Lastly, LASSO may arbitrarily select features from collinear groups, and further efforts should be put into interpreting the selected features and their odds ratios.

Conclusion

Serial HRV measurements over 2 hours can be used to improve prediction of 30-day IHM among septic patients in the ED compared to single time-point HRV measures taken at triage. This approach could be used in the ED to identify high-risk septic patients for appropriate management and disposition. In addition, the feature extraction and selection pipeline used in this paper could be applied to other short, noisy and incomplete time series data.