Efficient and Robust Heart Rate Estimation Approach for Noisy Wearable PPG Sensors Using Ideal Representation Learning

Photoplethysmography (PPG) is a non-invasive wearable sensing method used in millions of devices for heart rate monitoring. However, PPG signals are highly susceptible to a variety of noise sources, including motion artifacts, sensor noise, and biological factors, especially in real-world wearable settings. These make designing generalizable models to accurately interpret cardiac activities challenging. This paper proposes a focus shift from learning with noisy signals to utilizing the characteristics of a mathematically modelled PPG waveform in an adversarial setting to increase the signal-to-noise ratio. The results show the proposed approach is robust against noisy data. We evaluated the model in a user study (N=22), where it was tested against unseen PPG data collected from a new sensor and users under three different activity levels. Results showed the generalisability of the approach compared to the state-of-the-art and it maintains consistent performance improvements across diverse user activities. We successfully implemented our model on a commonly used (android) mobile device, confirming its ability to provide fast inferences in a resource-constrained setting.