Synthesis of Photoplethysmograph Signals from Noises with Desired Heart and Respiration Rates

Measurement of various kind of physiological parameters from the photoplethysmograph (PPG) signal has become an emerging area of research in recent years. Hence different state of the art deep learning algorithms are proposed in this domain. But the main problem of using such algorithms is scarcity of biomedical data of the target domain. Thus to alleviate such problems in this paper a novel method of synthesizing PPG data with required variance and desired heart rate (HR) and respiration rate (RR) is proposed. A deep CNN-LSTM autoencoder has been proposed which attempts to reduce an error having three components, targeting to match the morphology with desired variance with the template PPG, HR and RR with the required values. Over 600 samples were generated from the accelerometer noises and yielded a very good result in terms of HR, RR tracking, measured in terms of Mean absolute error (MAE) and KL Divergence, to find out the variance.