AI Outperforms Traditional Models at Predicting Heart Disease Risk in Chinese Adults
A transformer-based deep learning model predicts 10-year cardiovascular risk more accurately than Cox models and established risk scores in Chinese adults.
Summary
Researchers developed China-AIHeart, a sex-specific AI model using transformer deep learning architecture to predict 10-year cardiovascular disease risk in Chinese adults. Trained on nearly 157,000 participants and validated in two independent cohorts, the model outperformed traditional statistical approaches and existing risk calculators including China-PAR, PREVENT-ASCVD, and SCORE2 Asia-Pacific. It achieved C-statistics above 0.76 for both men and women, with strong calibration and net clinical benefit. The model comes in a full 22-predictor version and a simplified 15-predictor version, making it practical for real-world clinical use. This advance could meaningfully improve cardiovascular risk stratification for one of the world's largest populations.
Detailed Summary
Cardiovascular disease remains the leading cause of death globally, and accurate risk prediction is essential for early intervention. In Chinese populations, widely used Western risk calculators often perform poorly due to differences in genetics, lifestyle, and disease patterns. Traditional statistical models like Cox proportional hazards also have structural limitations when handling complex, non-linear interactions among risk factors.
Researchers from Shanghai Jiao Tong University and collaborating institutions across China developed China-AIHeart, a transformer-based deep learning model designed specifically for 10-year cardiovascular disease (CVD) risk prediction in Chinese adults. The derivation cohort comprised 156,790 participants from the China Cardiometabolic Disease and Cancer Cohort, all free of CVD at baseline, with a mean age of approximately 57 years. Separate models were built for men and women, using either 22 or 15 clinical predictors.
China-AIHeart demonstrated strong discrimination, with C-statistics of 0.767 in men and 0.780 in women — meaningfully better than Cox models using identical predictors (improvement of approximately 0.027–0.031). Net reclassification indices were substantial, at 0.478 in men and 0.560 in women, indicating the AI model correctly reassigned a significant proportion of patients to more accurate risk categories. External validation in the Xinjiang and CHARLS cohorts confirmed robust generalizability across diverse Chinese populations.
For clinicians, this means a practical, AI-powered tool could soon assist in identifying high-risk patients who might be missed by conventional scoring systems. The simplified 15-predictor version is particularly deployable in primary care settings where extensive testing is unavailable.
Caveats include that the summary is based on the abstract only, limiting assessment of predictor details, missing data handling, and follow-up duration. Additionally, findings apply specifically to Chinese adults and may not generalize to other ethnic populations without further validation.
Key Findings
- China-AIHeart achieved C-statistics of 0.767 (men) and 0.780 (women), outperforming all tested traditional risk scores.
- The AI model improved reclassification by 0.478 in men and 0.560 in women over Cox models with identical predictors.
- External validation in two independent Chinese cohorts confirmed robust performance, including a simplified 15-predictor version.
- Sex-specific models were developed separately for men and women, improving accuracy and clinical relevance.
- A simplified 15-predictor version maintained strong performance, making the tool viable for routine clinical settings.
Methodology
The study used a derivation cohort of 156,790 Chinese adults from the China Cardiometabolic Disease and Cancer Cohort, with external validation in the Xinjiang cohort and the CHARLS cohort. Transformer-based time-to-event prediction models were built with 22 and 15 predictors and compared against Cox models and established risk scores including China-PAR, PREVENT-ASCVD, and SCORE2 Asia-Pacific. Performance was assessed via C-statistic, calibration chi-square, Brier score, and net reclassification index.
Study Limitations
This summary is based on the abstract only, so details on predictor selection, missing data handling, follow-up duration, and model implementation are unavailable. The models were validated exclusively in Chinese populations and may not generalize to other ethnic groups without further study. As a retrospective development and validation study, prospective clinical impact trials are still needed.
Enjoyed this summary?
Get the latest longevity research delivered to your inbox every week.
Enter your email to subscribe: