Prediction of Heart Rate and Blood Oxygen from Physiological Signals

Wearable sensors have received massive interest from the research community due to their usage in a variety of applications. Particularly, great efforts have been oriented to non-invasive health monitoring of humans with the help of wearable sensors. For non-invasive health monitoring, a wide range of sensors is being developed and used. Among these sensors, some are more difficult to build and wear as compared to others. Blood Oxygen level and Heart Rate monitoring sensors, for instance, are among those sensors which are not easy to build and wear on the body continuously. In this paper, we present a novel idea of predicting blood oxygen level and heart rate from other physiological signals of easy to build and wear sensors like temperature, electrodermal activity, and acceleration. A state-of the-art supervised learning method called Random Forest is used to train and test the model on a publicly available dataset. The proposed method is also compared with several baseline regression techniques like KNN, and Support Vector regressor. Our proposed method achieved 0.9494 coefficient of determination (R2) and 3.26 root mean squared error for heart rate prediction. Whereas, a root mean squared error of 0.5589 and an R2 of 0.8565 for the prediction of blood oxygen was achieved.