A PROBABILISTIC APPROACH FOR HEART RATE VARIABILITY ANALYSIS USING EXPLICIT DURATION HIDDEN MARKOV MODELS

Monitoring of the temporal dynamics of the beat-to-beat intervals offers a non-invasive method for assessing autonomous nervous system activity. Recently it became feasible to continuously monitor cardiac activity through the pulse wave signal collected using wrist based sensors employing photoplethysmography (PPG). However, wearable sensor data collected in ambulatory setting is full of motion artifacts, baseline drift, and noise. New computational techniques are required to make reliable high level inferences from wearable sensor data. In this paper, we propose a probabilistic method for computing heart rate variability indices from noisy PPG sensor data collected in the natural environment. We model the joint distribution of beat labels and sensor data using an Explicit Duration Hidden Markov Model (EDHMM) and sample likely beat sequences from the posterior distribution conditioned on measured sensor data. Beat sequences produced by the EDHMM sampler can be used to calculate posterior distribution of arbitrary heart rate variability indices to form Bayesian estimates. Experimental validation with IEEE Signal Processing Cup data shows that our proposed framework can outperform state-of-the art methods in PPG signal analysis in continuous heart rate estimation.