Biosignal Fingerprinting: A Cross-Modal PPG-ECG Foundation Model
Researchers have introduced a novel cross-modal foundation model called the Multi-modal Masked Autoencoder (M2AE) to bridge the gap between diagnostic-rich ECG and ubiquitous wearable PPG signals for cardiovascular monitoring. Trained on over 3.4 million paired signals, M2AE generates compact latent representations termed biosignal fingerprints. These fingerprints uniquely encode an individual's cardiovascular state in a privacy-preserving, modality-agnostic format, allowing reuse across clinical tasks without exposing raw data or requiring retraining. The model demonstrates superior performance across seven downstream tasks, including cardiovascular disease classification, hypertension detection, and mortality prediction. Notably, it achieved an AUROC of 0.974 for five-class CVD classification and 0.877 for hypertension detection, marking a maximum improvement of 27.7% in AUROC compared to leading domain-specialist models. Crucially, the system maintains strong performance using only a single modality, facilitating deployment in resource-constrained, single-sensor environments typical of real-world wearable devices. This advancement has significant implications for continuous cardiovascular monitoring in both clinical and consumer health settings, offering a scalable solution to reduce global mortality from heart disease.
Wire timeline
Biosignal Fingerprinting: A Cross-Modal PPG-ECG Foundation Model
Researchers have introduced a novel cross-modal foundation model called the Multi-modal Masked Autoencoder (M2AE) to bridge the gap between diagnostic-rich ECG and ubiquitous wearable PPG signals for cardiovascular monitoring. Trained on over 3.4 million paired signals, M2AE generates compact latent representations termed biosignal fingerprints. These fingerprints uniquely encode an individual's cardiovascular state in a privacy-preserving, modality-agnostic format, allowing reuse across clinical tasks without exposing raw data or requiring retraining. The model demonstrates superior performance across seven downstream tasks, including cardiovascular disease classification, hypertension detection, and mortality prediction. Notably, it achieved an AUROC of 0.974 for five-class CVD classification and 0.877 for hypertension detection, marking a maximum improvement of 27.7% in AUROC compared to leading domain-specialist models. Crucially, the system maintains strong performance using only a single modality, facilitating deployment in resource-constrained, single-sensor environments typical of real-world wearable devices. This advancement has significant implications for continuous cardiovascular monitoring in both clinical and consumer health settings, offering a scalable solution to reduce global mortality from heart disease.
cs.AI updates on arXiv.org