Improving the quality of pulse rate variability derived from wearable devices using adaptive, spectrum and nonlinear filtering

• The aim is to analyze the filtering motion artifacts from PPG using the ACC signal. • The corrected PRVs were compared with reference HRV by calculating the RRMSE. • Filtering PPG signals with SF reduced the PRV error the most and NF the least. • In various laboratory experiments, similar results were achieved for SF + PP, SF + AF + PP, and NF + PP. • In signal affected by motion artifacts the SF + AF + NF + PP was the most beneficial. The popularity of wearable devices that record the photoplethysmographic (PPG) signal is increasing their use as monitors of circulatory and nervous systems function, including the extraction of pulse rate variability (PRV). Unfortunately, PPG is susceptible to interferences, including motion artifacts, which introduce errors into PRV. The aim of this study was to analyze the effectiveness of different approaches to filtering PPG (however, without recent deep learning methods) using an accelerometer (ACC) signal, in order to improve the accuracy of PRV extraction. This research concerns single and combined methods: adaptive filtering (AF) using the normalized least mean squares, spectrum filtering (SF) and nonlinear filtering (NF) using empirical mode decomposition. Filtering was complemented by postprocessing (PP): improving the accuracy of peak determination in PPG, correction of non-physiological values in PRV and its autoregressive modeling. The corrected PRVs were compared with reference heart rate variabilities (HRVs) extracted from the simultaneously recorded ECG by calculating the relative root mean square error (RRMSE). The results for different filtering sequences schemes indicate that filtering PPG signals with SF reduced the PRV error the most and NF the least. In various laboratory experiments, SF + PP, as well as SF + AF + NF + PP, resulted in the overall lower RRMSE, while in conditions similar to everyday life, the sequence of SF + AF + NF + PP proved to be the best sequence of methods (BS). The work shows that the use of the proposed BS of methods allows for improving the accuracy of PRV signal, increasing its similarity to the reference HRV for all types of analyzed experiments.