Pulse-by-pulse photoplethysmography quality index for signal reliability assessment based on pulse morphology

In the last decades, photoplethysmography (PPG) has been employed in a wide spectrum of applications, ranging from consumer devices to medical equipment. The pulsatile PPG signal recorded contains valuable information on the cardio-vascular and -respiratory system. Several features can be derived from the PPG and these can be used to infer physiological states of the human system, for instance to estimate the sleep structure or the fitness level of a subject [1]. As every sensing technology, the PPG can be corrupted by artifacts so it needs to be enhanced and, sometimes, parts of it have to be rejected. The rejection becomes fundamental especially when features are based on the shape of the signal. In literature, several methods are available to determine a quality index (QI) of parts of the PPG and use this QI as criterion for segment rejection. However, these methods exclude entire segments of the signal rather than single pulses, or they calculate a QI susceptible to the physiological PPG variation.
In this study, a new algorithm for single PPG pulse QI calculation is proposed. This QI ranges between 0 and 1 and it is assessed by comparing each pulse with a template. The template is derived from PPG by using dynamic time warping (DTW) barycenter averaging [2]. Each pulse is warped, using DTW, to maximize the match with the template and the QI is calculated as a normalized root mean square error of the remaining mismatch. The QI is resilient to physiological pulse deformations, but still able to quantify the pulse morphology corruption and to recognize artifacts.
The algorithm is validated on the Complex System Laboratory database [3], according to the ANSI/AAMI standards. For each pulse the beat location is calculated and it is rejected if QI is lower than 0.5. The positive predictive value and sensitivity (PPV, SEN) are calculated with respect to human beat annotations, with a true positive detection criterion of 30 ms distance from the annotated beat. The developed algorithm has a PPV of 99.43% and a SEN of 95.43% while the one in [3] has, respectively, 97.98% and 98.99%. The QI thresholding allows to obtain a significantly higher PPV, a consequence of an improved corrupted beat rejection, at the expense of a lower sensitivity. The proposed algorithm provides a single pulse QI, resilient to physiological deformation, based on the morphology comparison with a data derived template. The algorithm gives the possibility to choose the QI depending on the features that will be derived. For instance, in case of morphology related features, a high QI threshold can increase the likelihood that only uncorrupted beats are kept.