Multi-Task Learning for Photoplethysmographic Measurement of Heart Rate

Abstract

Autonomous vehicles and telemedine are both applications that require remote monitoring of vital signs. In recent years, research has focused on photoplethysmographic methods because they are cheap to implement and can be used remotely. However, robustness of these techniques within "real-world" environments needs to improved before they can be widely deployed. In this work we assemble a larger and more diverse dataset than those publicly available. We develop eight models on this dataset, designed to return improved results by adjusting for skin-tone. These models can be broken into three classes: Purely Photoplethysmographic, Single-task Learning, and Multi-task Learning. While previous research has focused almost entirely on the first class of models, we have found that the last two classes, particularly Multi-task learning, yield significantly more accurate results.