A Deep Learning Approach to Estimate SpO2 from PPG Signals

Abstract

Blood oxygen saturation level (SpO2) is one of the vital parameters determining the hemostability of a patient, besides heart rate (HR), respiratory rate (RR) and blood preasure (BP). In emergency situations with a high number of injured persons, during the second triage until arrival to a medical facility, continuously following the SpO2 level in real time is of outmost importance. Using a smart patch-like device attached to a injured’s chest that contains a Photoplethysmogram (PPG) sensor, one can obtain the SpO2 parameter. Our interest in the process of the smart patch prototype development is to investigate the monitoring of a blood oxygen saturation level by using the embedded PPG sensor. We explore acquiring the SpO2 by extracting the set of features from the PPG signal utilizing Python toolkit HeartPy in order to model a Deep neural network regressor. The PPG signal is preprocessed by various filtering techniques to remove low/high frequency noise. The model was trained and tested using the clinical data collected from 52 subjects with SpO2 levels varying from 83 - 100%. The best experimental results considering the SpO2 interval [83,95) were achieved with a PPG signal of 10 seconds length (MAPE 2.00% and 7.21% of big errors defined as absolute percentage errors (APE) equal or greater than 5).