Exploring Self-supervised Learning for PPG-based Heart-Rate Estimation

In recent years, Heart Rate (HR) monitoring is becoming increasingly widespread in wrist-worn devices where low-cost photoplethysmography (PPG) sensors are installed. On the other hand, the accuracy of PPG-based HR tracking is often compromised by Motion Artifacts (MAs), which result from movements of the subject’s arm, and cause degradation in the quality of the PPG signal gathered. To mitigate this issue, the PPG signal is commonly combined with acceleration measurements obtained from an inertial sensor. In the state-of-the-art, many traditional methods based on temporal and frequency filters and, more recently, deep learning algorithms have been exploited to combinate the information from these two sensors. In this thesis, driven by the recent achievements of self-supervised learning, we investigate its application for PPG-based heart rate tracking. The main disadvantage of using this kind of approach is that it requires a large amount of data to process. This problem has only been overcome in recent years, with the introduction of large datasets such as PPG-Dalia and WESAD. Once data have been made available, self-supervised learning could represent an alternative and innovative solution to obtain a good level of accuracy in model predictions, allowing better generalization. Adopting Masked Autoencoders as reference model, we exploit self-supervised learning to teach a neural network how to reconstruct PPG signals both in time and in frequency. During this pre-training step, the layers of the network learn to extract good features for the task. Subsequently, we fine-tune our Masked Autoencoder, substituting the decoder with a regressive tail and leveraging the knowledge acquired from the previous phase to estimate the heart rates of the patients. We reach an average of Mean Absolute Error (MAE) over all patients of 6.16 Beats Per Minute (BPM) for PPG-Dalia and 5.42 BPM for WESAD (better than state-of-the-art of 2.05 BPM). Applying an additional post-processing step, the MAE is further reduced to 5.81 BPM and 4.95 BPM, respectively. Furthermore, we analyze the advantages of using Transfer Learning during the self-supervised pre-training step. This involves pre-training our reference model on one dataset and then fine-tuning it on the other, and vice versa. Through this approach, we demonstrate that, for the majority of the patients, it is possible to improve performance and further decrease MAE.