Robust probabilistic modeling of photoplethysmography signals with application to the classification of premature beats

In this paper we propose a robust approach to model photoplethysmography (PPG) signals. After decomposing the signal into two components, we focus the analysis on the pulsatile part, related to cardiac information. The goal is to enable a deeper understanding of the information contained in the pulse shape, together with that derived from the rhythm. Our approach combines functional data analysis with a state space representation and guarantees fitting robustness and flexibility on stationary signals, without imposing a priori information on the waveform and heart rhythm. With a Bayesian approach, we learn the distribution of the parameters, used for understanding and monitoring PPG signals. The model can be used for data compression, for inferring medical parameters and to understand condition-related waveform characteristics. In particular, we detail a procedure for the detection of premature contractions based on the residuals of the fit. This method can handle both atrial and ventricular premature contractions, and classify the type by only using information from the model fit.