Prediction of Noise-Type and its Elimination from Photoplethysmograph Signal

Non-invasive measurement of photoplethysmograph (PPG) signal is an emerging research area. However the presence of noises in such signals may give erroneous interpretations. This paper thus proposes a technique to identify the noise model in the PPG signal, followed by the suitable de-noising techniques. Reservoir computing (RC) model has been used for the classification purpose which reduces numbers of learnable parameters and provided fast training. Cepstrum analysis has been applied for the elimination of convolutive noises. The method has been tested on 700 PPG windows and yielded good results from the point of view of PPG de-noising and estimation of respiration rate (RR) in terms of NRMSE and PCC with a standalone hardware implementation on a low-cost resource-constrained embedded device (Raspberry Pi Zero W) having an end-to-end latency and memory engagement of 165.34 s and 331.10 KB respectively.