Ultra-Short Term Heart Rate Variability Estimation Using PPG and End-to-End Deep Learning

Heart rate variability (HRV) is a marker that could be indicative of critical health events. Continuous monitoring of HRV in intensive care unit (ICU) holds great promise for improving patient's prognosis. However, current methods requiring measuring HRV using electrocardiogram (ECG) sensors over an extended duration offer limited applicability for continuous and disturbance-free HRV tracking in ICU cases. This underscores the need for new methods that can rapidly measure HRV from a single unobtrusive sensor. Here, we explore the feasibility of using an end-to-end deep learning model to enable real-time ultra-short term HRV (usHRV) estimation from 10-s segments of photoplethysmogram (PPG) signal, acquired from an optical sensor commonly present in ICU settings. Validated on testing data from hold-out ICU subjects, the results from our deep learning model are in excellent agreement with the ECG-based references for estimating usHRV metrics, achieving R-squared scores of 1.00, 0.89 and 0.93 in estimating MeanNN, SDNN and RMSSD, respectively. Furthermore, our model outperforms PPG-based manually-extracted pulse rate variability (PRV), an HRV surrogate, in both accuracy and robustness. Overall, our study offers a promising perspective for the use of small-scale deep learning models in achieving reliable usHRV tracking in intensive care scenarios.