A Random Forest machine learning model incorporating multiple lipid markers significantly outperformed the Framingham Risk Score in predicting sudden cardiac death (AUC 0.824 vs 0.737).
Cohort (n=3,172)
Sí
Does a machine learning model incorporating multiple lipid markers improve the prediction of sudden cardiac death in a community-based population compared to the Framingham Risk Score?
A Random Forest machine learning model incorporating lipid markers accurately predicts long-term sudden cardiac death risk in a community cohort, outperforming the Framingham Risk Score.
Estimación del efecto: ΔAUC 0.087
Tasa de eventos absoluta: 0.824% vs 0.737%
valor p: p=0.027
Abstract Background Sudden cardiac death (SCD) is a major contributor to cardiovascular mortality, but reliable long-term risk prediction in community-based populations remains limited. Machine learning (ML) offers potential advantages, yet its application to SCD prediction remains comparatively limited, particularly in community-based populations. Methods We used data from the Chin-Shan Community Cardiovascular Cohort (CCCC), a prospective community-based cohort in Taiwan enrolling adults aged ≥ 35 years from 14 villages. Participants were geographically partitioned into training/internal validation and independent external validation cohorts. After feature preselection using the Boruta algorithm, six ML models were developed (Light Gradient Boosting Machine, Random Forest RF, Logistic Regression, Support Vector Machine, Multilayer Perceptron, and K-Nearest Neighbours), with class imbalance addressed using appropriate techniques. Model performance was evaluated using discrimination and classification metrics, including the area under the curve (AUC), positive predictive value, and negative predictive value (NPV). The optimal model was interpreted using SHapley Additive exPlanations and implemented as a web-based risk calculator. Results A total of 3,172 participants were included (median age IQR: 54.9 45.8–64.0 years; 47.4% male), with 74 SCD events observed over a median follow-up of 15.9 years (IQR, 13.1–16.9 years). Ten non-collinear predictors were preselected, and six ML models were developed and validated. The RF showed the highest discrimination in internal (AUC: 0.824) and external validation (AUC: 0.815) and was the only model to significantly outperform the Framingham Risk Score. The RF demonstrated consistently high NPVs (> 99%) in both internal and external validation cohorts and was implemented as a web-based risk calculator. Conclusions We developed and externally validated a RF model for long-term SCD risk prediction with high NPV, supporting its potential utility for identifying low-risk individuals in community settings, pending further validation. The model has been implemented as an online risk calculator, with further validation in larger and diverse populations warranted.
Chen et al. (Sat,) conducted a cohort in Sudden cardiac death (n=3,172). Random Forest machine learning model vs. Framingham Risk Score was evaluated on Discrimination of sudden cardiac death risk (AUC) in internal validation (ΔAUC 0.087, p=0.027). A Random Forest machine learning model incorporating multiple lipid markers significantly outperformed the Framingham Risk Score in predicting sudden cardiac death (AUC 0.824 vs 0.737).