Cardiovascular risk stratification with heart rate topics

Recent work on heart rate motifs (HRM) has demonstrated that information in short heart rate patterns may be useful in identifying patients at elevated risk of cardiovascular death (CVD) following acute coronary syndrome. The information in HRM complements a variety of other clinical metrics including electrocardiographic (ECG) measures. While the HRM approach has value, it suffers from a focus on identifying and using information related to only a small number of discriminative patterns in heart rate time series, which loses valuable information among the full set of patterns in the data. We present a method based on topic models, an approach traditionally used in text analysis, to learn structure in the full set of short heart rate patterns in long-term ECG recordings. This model provides an interpretable representation of long-term ECG recordings and finds relationships amongst all short heart rate patterns across the entire patient population. When evaluated on data from 4,557 patients admitted with non-ST-elevation acute coronary syndrome, we show that heart rate topic models significantly improve risk stratification. This improvement is consistent even when considering information already available through the TIMI risk score and left ventricular ejection fraction, as well as several heart rate variability metrics.