Multimodal cardiovascular risk profiling using self-supervised learning of polysomnography.

Area of Science:

• Biomedical Engineering
• Cardiology
• Sleep Medicine

Background:

• Polysomnography (PSG) is standard for sleep disorder diagnosis but underutilized for predicting future health risks.
• Cardiovascular disease (CVD) risk assessment traditionally relies on factors like age, blood pressure, and cholesterol.
• Exploring novel data sources like PSG signals for enhanced CVD risk prediction is crucial.

Purpose of the Study:

• To develop and validate an interpretable framework using self-supervised learning on PSG data for CVD risk prediction.
• To identify physiological patterns in Electroencephalography (EEG), Electrocardiography (ECG), and respiratory signals linked to CVD outcomes.
• To assess the framework's performance independently of manual sleep stage annotations.

Main Methods:

• A self-supervised deep learning model was created to extract patterns from multi-modal PSG signals (EEG, ECG, respiratory).
• The model was trained on 4,398 participants, generating projection scores contrasting individuals with and without CVD.
• External validation was performed on an independent cohort of 1,093 participants.

Main Results:

• Projection scores from ECG, EEG, and respiratory signals showed distinct, clinically meaningful patterns predictive of CVD outcomes.
• ECG features predicted cardiac conditions and CVD mortality; EEG features predicted hypertension and CVD mortality.
• Integrating projection scores with the Framingham Risk Score significantly improved CVD risk prediction (AUC 0.607-0.965 in internal, 0.710-0.807 in external validation).

Conclusions:

• The developed framework generates individualized CVD risk scores directly from PSG data, offering a novel risk stratification tool.
• These interpretable projection scores can enhance clinical risk assessment and personalized cardiovascular care.
• The study demonstrates the significant, untapped potential of PSG signals for predicting cardiovascular health outcomes.