Self-Supervised Autoencoder Network for Robust Heart Rate Extraction from Noisy Photoplethysmogram: Applying Blind Source Separation to Biosignal Analysis

TL;DR

The study tackles robust heart rate extraction from noisy PPG by framing the problem as blind source separation and employing a self-supervised multi-encoder autoencoder (MEAE). Trained on a large polysomnography dataset (MESA) without preprocessing, the MEAE isolates heartbeat-related sources and is then applied to noisy daily-life PPG, outperforming ICA, NMF, and BRDAE baselines. On a 9-subject daily-activity dataset, the selected MEAE source achieves about 4.9 bpm RMSE and a 0.74 Pearson correlation with ECG, marking a substantial improvement over raw PPG. The findings demonstrate the potential of BSS with MEAE for robust biosignal analysis in real-world noisy conditions, while outlining limitations and avenues for future work.

Abstract

Biosignals can be viewed as mixtures measuring particular physiological events, and blind source separation (BSS) aims to extract underlying source signals from mixtures. This paper proposes a self-supervised multi-encoder autoencoder (MEAE) to separate heartbeat-related source signals from photoplethysmogram (PPG), enhancing heart rate (HR) detection in noisy PPG data. The MEAE is trained on PPG signals from a large open polysomnography database without any pre-processing or data selection. The trained network is then applied to a noisy PPG dataset collected during the daily activities of nine subjects. The extracted heartbeat-related source signal significantly improves HR detection as compared to the original PPG. The absence of pre-processing and the self-supervised nature of the proposed method, combined with its strong performance, highlight the potential of MEAE for BSS in biosignal analysis.

Figures (6)

• Figure 1: Diagram outlining the architecture of the proposed multi-encoder autoencoder used in this study to extract heart-related source signal from photoplethysmogram. Details of the network and the implementation for training can be found in our GitHub repository.
• Figure 2: Diagram showing the inference step for the proposed method. An encoder is selected, and the output from all other encoders are masked with zeros. Then the encodings are passed to the decoder, yielding a source prediction corresponding to the selected encoder. Details of the network and the implementation for training can be found in our GitHub repository. $\hat{S}_n$ is the nth predicted source corresponding to the nth encoder, $E_n$.
• Figure 3: Detection of heart rate from electrocardiogram (top), original photoplethysmogram (middle), and the first source signal (bottom) generated from the optimized multi-encoder autoencoder network. The detected peaks corresponding to each signals are used to generate beat-by-beat heart rates for each cardiac cycle.
• Figure 4: Electrocardiogram (top row), photoplethysmogram (2nd row), source signals (3rd–10th rows) from the optimized multi-encoder autoencoder, and PPG reconstruction (bottom row). 48-second samples (clipped on both sides to about 46 seconds to remove noisy edges) from the MESA dataset (a) and the noisy PPG dataset (b) are shown. Dashed vertical lines indicate heartbeat peaks detected from the ECG.
• Figure 5: Bland-Altman plots comparing heart rate detection from the original photoplethysmogram (a) and the selected source signal (b) against the reference electrocardiogram.