Quantitative analysis of smartphone PPG data for heart monitoring

The field of app-based PPG monitoring of cardiac activity is promising, yet classification of heart rhythms in normal sinus rhythm (NSR) or atrial fibrillation (Afib) is difficult in the case of noisy measurements. In this work, we aim at characterizing a dataset of 1572 subjects, whose signals have been crowdsourced by collecting measurements via a dedicated smartphone app, using the embedded camera. We evaluate the distributions of three features of our signals: the peak area, amplitude and the time interval between two successive pulses. We evaluate if some factors affected the distributions, discovering that the strongest effects are for age and BMI groupings. We evaluate the results agreement between the R G B channels of acquisition, finding good agreement between the first two. After finding signal quality indexes in literature, we use a subset of them in a classification task, combined with dynamic time warping distance, a technique that matches a signal to a template. We achieve an accuracy of 89% on the test set, for binary quality classification. On the chaotic temporal series we evaluate the appearance of different types of rhythms on Poincaré plots and we quantify the results by a measure of their 3D spread. We perform this on a set of 20 subjects, 10 NSR and 10 Afib, finding significant differences between their 3D morphologies. We extend our analysis to the larger dataset, obtaining some significant results.