Abstract
Wrist based devices, like smart-watches, fitness bands and health monitors all provide a common sensor called Photoplethysmography (PPG) to measure optical pulse signal. This is usually used to derive the instantaneous heart-rate (HR), which is useful while doing any exercise or to monitor on a regular basis for chronic patients. However, one major issue with the signal is that it is easily corrupted by ambulatory motion generated by hand movements of the subject. Since, these devices also come equipped with an independent motion sensor, namely a tri-axes accelerometer, researchers have taken interest in trying to correct the motion artifact in PPG using the accelerometer as a reference noise signal. However, it is not a trivial problem and hence, even after a substantial body of research, the problem remains unsolved, especially when considering on-premise estimation due to the resource-constrained nature of wearable devices. In this paper, we aim to solve this problem using subspace based learning approach. Though this approach has been utilized before, we have added some novel steps to the algorithm pipeline and also made modifications so that the algorithm can be possibly run on a typical wearable device. Our preliminary results show efficacy and promise of our proposed approach.