EnhancePPG: Improving PPG-based Heart Rate Estimation with Self-Supervision and Augmentation

TL;DR

EnhancePPG tackles motion-artifact robust HR estimation from PPG by combining self-supervised pre-training with data augmentation, enabling learning from large unlabeled datasets. The method retools the PULSE architecture into an autoencoder for SSL pre-training and applies DA during pre-training with unlabeled data, followed by LOSO-based fine-tuning on labeled PPG-DaLiA data. It achieves a new state-of-the-art MAE of 3.54 BPM on PPG-DaLiA (a 12.2% improvement) with negligible latency increase, and shows pronounced improvements for critical subjects, while remaining deployable on low-power MCU hardware. The work demonstrates that SSL and DA can leverage unlabeled data to bolster wearable HR estimation under motion artifacts, with broad applicability to other time-series models.

Abstract

Heart rate (HR) estimation from photoplethysmography (PPG) signals is a key feature of modern wearable devices for health and wellness monitoring. While deep learning models show promise, their performance relies on the availability of large datasets. We present EnhancePPG, a method that enhances state-of-the-art models by integrating self-supervised learning with data augmentation (DA). Our approach combines self-supervised pre-training with DA, allowing the model to learn more generalizable features, without needing more labelled data. Inspired by a U-Net-like autoencoder architecture, we utilize unsupervised PPG signal reconstruction, taking advantage of large amounts of unlabeled data during the pre-training phase combined with data augmentation, to improve state-of-the-art models' performance. Thanks to our approach and minimal modification to the state-of-the-art model, we improve the best HR estimation by 12.2%, lowering from 4.03 Beats-Per-Minute (BPM) to 3.54 BPM the error on PPG-DaLiA. Importantly, our EnhancePPG approach focuses exclusively on the training of the selected deep learning model, without significantly increasing its inference latency

Figures (3)

• Figure 1: a) Modified PULSE architecture used in fine-tuning; b) our PULSE-autoencoder architecture used in pre-training
• Figure 2: MAE results on PPG-DaLia considering architectural modifications ("Mod. PULSE") and our EnhancePPG ("Pre-train" and Augmentation).
• Figure 3: MAE results on PPG-DaLiA on best phase to apply DA.