Predicting cardiovascular disease risk using photoplethysmography and deep learning.

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence in Healthcare

Background:

• Cardiovascular diseases (CVDs) are a leading cause of premature death, particularly in low- and middle-income countries.
• Existing CVD risk assessment tools often require physical examinations or lab tests, limiting accessibility in resource-constrained settings.
• Photoplethysmography (PPG) sensing, available on smartphones, presents a potential low-cost solution for widespread CVD screening.

Purpose of the Study:

• To investigate the efficacy of a deep learning-based CVD risk score (DLS) utilizing PPG signals for predicting major adverse cardiovascular events (MACE).
• To compare the performance of the PPG-based DLS against a refitted WHO risk score.
• To assess the potential of PPG-based risk prediction for primary prevention in resource-limited areas.

Main Methods:

• Developed a deep learning model (DLS) to predict 10-year MACE risk using age, sex, smoking status, and PPG data.
• Compared DLS performance against an office-based refit-WHO score using UK Biobank data (development and test sets).
• Evaluated model performance using C-statistic for discrimination and mean absolute calibration error.

Main Results:

• The DLS achieved a C-statistic of 71.1%, demonstrating non-inferiority to the refit-WHO score (70.9%) in the test dataset.
• The DLS showed satisfactory calibration with a 1.8% mean absolute calibration error.
• Incorporating DLS features improved the C-statistic of the office-based score by 1.0%.

Conclusions:

• A deep learning PPG-based CVD risk score can predict 10-year MACE risk comparably to traditional office-based scores.
• PPG-derived features, independent of heart rate, contribute to risk prediction, suggesting a novel approach to CVD assessment.
• This study supports the potential of PPG-based strategies for accessible, community-based primary CVD prevention, especially in resource-limited settings.