Estimating Short-Term Heart Rate Variability from Ultra-Short PPG: Mitigating Systematic Bias Using Deep Learning

Estimating standard short-term (ST, 5 min) heart rate variability (HRV) metrics (derived from electrocardiogram (ECG)) using ultra-short-term (UST, $< 5$ min) segments of photoplethysmogram (PPG) is desirable for real-time, cost-effective HRV monitoring in critical care settings. Traditional PPG-based HRV estimation relies on pulse rate variability (PRV), which is limited in accuracy due to physiological differences between PPG and ECG signals. Here, we propose a data-driven, end-to-end deep learning approach that estimates ECG-based ST HRV metrics directly from UST PPG segments (as short as 10 s). The key idea is to leverage deep learning to characterize full aspects of the PPG waveform, for learning domain-invariant HRV information shared by short PPG segments and longer ECG recordings. It is shown that the consistent and systematic shift in bias observed in PRV-based HRV estimates, can be mitigated through supervised deep learning using ground-truth ST HRV metrics. HRV estimates from deep learning and PRV-based baselines are compared across 7 ST and UST PPG durations using the VitalDB dataset (4,759 subjects). Models trained on VitalDB are further validated on the BIDMC dataset (24 subjects) to evaluate their generalizability. On the VitalDB dataset, deep learning-based HRV estimates consistently outperform PRV-based baselines as surrogates of ECG-derived ST HRV metrics across time, frequency and nonlinear domains, achieving high correlations ($\rho >0.7$) and low biases ($|\text{Cohen's } \mathrm{d}|< 0.2$) using UST PPG segments of $\geq 120$ s duration. Meanwhile, models trained on the VitalDB dataset achieve moderate to high correlation ($\rho >0.5$) and low to moderate bias ($|\mathrm{d}|< 0.8$) for estimating the majority of ST HRV metrics from $\geq 30$ s PPG segments in the BIDMC dataset, showcasing promising cross-dataset generalizability. This study demonstrates the feasibility of using deep learning to infer clinically relevant long-term cardiovascular information from short physiological signals for predictive health monitoring.