[Paper Review] SiamAF: Learning Shared Information from ECG and PPG Signals for Robust Atrial Fibrillation Detection

Zhicheng Guo, Cheng Ding+5 more|arXiv (Cornell University)|Oct 13, 2023

ECG Monitoring and AnalysisMedicine3 citations

TL;DR

SiamAF proposes a Siamese deep learning framework that jointly learns shared features from ECG and PPG signals to improve robust atrial fibrillation (AF) detection. By leveraging a contrastive loss to align modality-specific representations around peak features, the model achieves state-of-the-art performance on multiple external datasets and generalizes well even with only 1% of labeled data.

ABSTRACT

Atrial fibrillation (AF) is the most common type of cardiac arrhythmia. It is associated with an increased risk of stroke, heart failure, and other cardiovascular complications, but can be clinically silent. Passive AF monitoring with wearables may help reduce adverse clinical outcomes related to AF. Detecting AF in noisy wearable data poses a significant challenge, leading to the emergence of various deep learning techniques. Previous deep learning models learn from a single modality, either electrocardiogram (ECG) or photoplethysmography (PPG) signals. However, deep learning models often struggle to learn generalizable features and rely on features that are more susceptible to corruption from noise, leading to sub-optimal performances in certain scenarios, especially with low-quality signals. Given the increasing availability of ECG and PPG signal pairs from wearables and bedside monitors, we propose a new approach, SiamAF, leveraging a novel Siamese network architecture and joint learning loss function to learn shared information from both ECG and PPG signals. At inference time, the proposed model is able to predict AF from either PPG or ECG and outperforms baseline methods on three external test sets. It learns medically relevant features as a result of our novel architecture design. The proposed model also achieves comparable performance to traditional learning regimes while requiring much fewer training labels, providing a potential approach to reduce future reliance on manual labeling.

Motivation & Objective

• To address the challenge of noisy, low-quality wearable ECG and PPG signals in AF detection.
• To improve model generalization across diverse recording settings and hardware by leveraging complementary information from both ECG and PPG.
• To reduce reliance on large-scale manual labeling by enabling effective semi-supervised learning through joint modality alignment.
• To develop a unified model that performs well on both ECG and PPG inputs without performance trade-offs.

Proposed method

• A Siamese neural network architecture is used to process ECG and PPG signals in parallel, sharing the same backbone (ResNet) and projection head.
• A novel contrastive loss function encourages the model to learn shared representations between ECG and PPG signals, particularly focusing on R-peak features.
• The model is trained using a joint learning objective that combines classification loss (for AF prediction) and contrastive loss (for modality alignment).
• The framework enables semi-supervised learning by leveraging unlabeled data through the contrastive loss, improving sample efficiency.
• Peak-only signal representations are used to verify that the model primarily relies on R-peak information for classification.
• Dimensionality reduction and feature visualization are used to interpret and validate the medically relevant features learned by the model.

Experimental results

Research questions

• RQ1 Can joint learning from ECG and PPG signals improve the robustness and generalization of AF detection models in real-world noisy settings?
• RQ2 Does learning shared features—particularly R-peak patterns—between ECG and PPG lead to better performance than single-modality models?
• RQ3 To what extent can a contrastive learning objective reduce dependency on large amounts of labeled data in AF detection?
• RQ4 Can a single unified model achieve high performance on both ECG and PPG inputs without modality-specific fine-tuning?
• RQ5 What types of physiological features (e.g., peak morphology, P-wave presence) does the model actually learn to rely on for classification?

Key findings

• SiamAF outperforms multiple state-of-the-art baselines, including DeepBeat and single-modality CNNs, on three independent external test sets with diverse data sources and recording conditions.
• The model achieves comparable performance to fully supervised models while using only 1% of the training labels, demonstrating strong sample efficiency.
• Feature visualization confirms that the model learns medically relevant, shared representations centered on R-peak morphology across ECG and PPG signals.
• The model generalizes well to out-of-distribution data, such as ambulatory PPG from the Simband dataset and hospital ECG from Institution B, outperforming models trained on single modalities.
• Peak-only signal projections show high overlap between ECG and PPG representations, indicating that R-peak features are the primary shared signal for AF detection.
• The contrastive loss significantly improves model performance and generalization, even when labels are scarce, validating its role in guiding feature learning toward robust, shared patterns.