Development and External Validation of a Deep Learning Algorithm for Prognostication of Cardiovascular Outcomes

Area of Science:

• Medical Informatics
• Cardiovascular Medicine
• Machine Learning

Background:

• Cardiovascular disease (CVD) remains a leading cause of mortality globally.
• Accurate risk prediction is crucial for timely intervention and prevention strategies.
• Existing models may not fully leverage longitudinal data for enhanced discrimination.

Purpose of the Study:

• To evaluate the enhanced discriminative accuracy of a deep learning (DL) algorithm for cardiovascular disease (CVD) risk prediction.
• To compare the performance of a DL model against traditional Cox hazard regression using repeated-measures data.
• To assess the generalizability of the DL model across different populations.

Main Methods:

• Development of two CVD prediction models: a Cox regression model and a DL model, utilizing the National Health Insurance Service-Health Screening Cohort (NHIS-HEALS).
• Internal validation of both models within the NHIS-HEALS cohort.
• External validation in two independent cohorts: the National Health Insurance Service-National Sample Cohort (NHIS-NSC) in Koreans and the Rotterdam Study in Europeans.
• Inclusion of a large sample size: 412,030 adults in NHIS-HEALS, 178,875 in NHIS-NSC, and 4,296 in the Rotterdam Study.

Main Results:

• The DL model consistently outperformed Cox regression in both internal and external validation cohorts across different populations.
• In internal validation (NHIS-HEALS), DL achieved higher C-statistics (0.896 men, 0.921 women) and significant improvements in reclassification (NRI: 24.8% men, 29.0% women).
• External validation in NHIS-NSC and the Rotterdam Study also showed superior performance of the DL model, with notable improvements in C-statistics and net reclassification indices.

Conclusions:

• A deep learning algorithm demonstrates superior discriminative accuracy for cardiovascular disease risk prediction compared to Cox model approaches.
• The DL model's ability to utilize repeated-measures data enhances its predictive power.
• These findings suggest that DL algorithms hold significant promise for improving CVD risk stratification and personalized medicine.