Enhancing PPG-Based Heart Rate Estimation combining Data Augmentation and Model Pre-training

Accurate heart rate (HR) estimation from photoplethysmographic (PPG) signals is critical for health monitoring applications, such as cardiovascular disease management, fitness and activity tracking, and detecting arrhythmia or other heart-related abnormalities. Without subjects’ movements and with personalized subject data, nowadays HR- estimation can be considered a “solved problem”, even with low-cost commercial smartwatches. However, in a more general and realistic scenario of daily life, this task remains challenging due to motion artefacts (MA) and inter-subject variability. This thesis addresses these challenges by proposing a series of improvements to existing neural network architectures for more robust and accurate HR estimation. We first explored different Deep Learning approaches, starting with a Masked Autoencoder for unsupervised feature learning. While this model provided an initial reference, it did not meet the desired performance in HR estimation. Subsequently, we focused on improving the state-of-the-art PULSE (Ppg and imU signal fusion for heart rate Estimation) architecture, a model designed specifically for HR estimation from PPG signals. To better capture localized features, we replaced the dilated convolutions in the original model with standard convolutions, increasing the filter size and eliminating the need for dilation. Building on our new architecture, we introduced a pre-training step where the model task is the reconstruction of PPG signals. To implement this, we restructured the PULSE model into an encoder-decoder architecture. Specifically, we replaced the final two linear layers with a decoder that mirrors the encoder architecture, utilizing both the multi-headed attention mechanism and transposed convolutional layers to reconstruct the input signal from the latent space. Inspired by U-Net, we also introduced two skip connections between the corresponding encoder and decoder layers to retain spatial information during the up-sampling process. As a last step, to further improve model robustness and generalization, we applied a variety of data augmentation (DA) techniques to expand the training data, particularly targeting heart rate variability. These augmentations, inspired by a recent paper on improving PPG-based HR monitoring with synthetically generated data, proved highly effective in enhancing model performance, especially for subjects with atypical beats-per-minute (BPM) ranges. In our experiments, we applied the pre-training on the PULSE model using data from the public WESAD dataset and from a new un-labelled dataset coming from West Attica University, to form a comprehensive dataset of 490,000 samples. The model has been finally evaluated on the PPG-Dalia dataset, the widest labelled dataset to compare with state-of-the-art models, using the Leave-One-Subject-Out (LOSO) protocol. In conclusion, our proposed enhancements to the PULSE model led to a 12.4% reduction in MAE on the PPG-DaLiA dataset, achieving a final MAE of 3.53 compared to best state-of-the-art competitor, PULSE, that achieves 4.03 BPM of MAE with identical model complexity. Note that all our enhancements are applied during training, therefore not modifying the complexity of the model inference, usually executed on a low-power edge device such as a smartwatch. This work sets a new benchmark in HR estimation accuracy from PPG signals, demonstrating the power of pre-training and data augmentation techniques in improving the performance of an already state-of-the-art architecture.