Foundation models for cardiovascular disease detection via biosignals from digital stethoscopes

Area of Science:

• Cardiology
• Artificial Intelligence
• Biomedical Signal Processing

Background:

• Auscultation and electrocardiograms (ECG) are key cardiac exam components.
• Digital stethoscopes now enable simultaneous high-quality acoustic and ECG signal acquisition.

Purpose of the Study:

• To develop foundation models for cardiovascular disease detection using synchronously captured phonocardiogram (PCG) and ECG data.
• To leverage self-supervised learning on large unlabeled datasets for robust model training.

Main Methods:

• Trained foundation models on PCG and ECG data from digital stethoscopes in clinical practice.
• Utilized a masked autoencoder framework extended for multiple synchronous signals.
• Fine-tuned pre-trained models for various cardiovascular disease detection tasks.

Main Results:

• Demonstrated the effectiveness of foundation models for cardiovascular disease detection.
• Achieved superior performance with large-capacity models, overcoming limitations of small labeled datasets.
• This is the first study to build foundation models for synchronously captured PCG and ECG data.

Conclusions:

• Foundation models pre-trained on unlabeled synchronous PCG-ECG data offer a powerful approach for cardiovascular disease detection.
• Self-supervised learning with foundation models can enhance diagnostic capabilities, especially when labeled data is scarce.
• This methodology enables the use of advanced AI models for improved cardiac diagnostics.