ABSTRACT

Introduction: Cardiovascular disease (CVD) has been the leading cause of death in Malaysia since the 1980s, with myocardial infarction (MI) and stroke being significant contributors.

Research gao: Conventional risk scores like the Framingham Risk Score (FRS) are widely used for predicting CVD, but their 10-year risk assessment often lacks population specificity, particularly for the Malaysian demographic, and fails to capture short- to medium-term risks. This limitation may lead to missed opportunities for timely interventions, especially given the rapid progression of CVD in some patients.

Hypothesis: In response to this gap, this study developed a 3-year CVD risk prediction and event classification model tailored to Malaysia's diverse demographic landscape, addressing the need for a more accurate and timely risk assessment tool.

Data: Utilizing data from the REDISCOVER registry, which included 12,617 participants from eight Malaysian states, two distinct models were constructed. The first model predicted the occurrence of CVD events (CVD vs. No CVD), while the second classified specific CVD events, particularly MI and stroke.

Methodology: Data preprocessing involved median imputation for missing values and resampling to balance the classes. Feature selection was performed using Recursive Feature Elimination(RFE), SHAP-based selection, backward elimination as well as correlation analysis for noise reduction. The models trained included Random Forest (RF), XGBoost, Decision Trees (DT), Gradient Boosting Machine (GBM), Support Vector Classifier (SVC), and K- Nearest Neighbors (KNN). Ensemble models were also created using a stacking approach. Model performance was evaluated based on accuracy, area under the curve (AUC), classification reports, and confusion matrices.

Results: In the first model, the RF model using the top 26 selected features achieved an AUC of 0.8587, outperforming the Framingham Risk Score (FRS), which had an AUC of 0.8343. The RF model also outperformed FRS in terms of the Net Reclassification Index (NRI), showing a 19% improvement at a 5% cut-off point. While the calibration of the RF model was similar to FRS, the Brier score was slightly higher for the RF model (0.01176) compared to FRS (0.01173), and the log loss was marginally greater (0.05720 for RF vs. 0.0568 for FRS). The second model, aimed at classifying specific CVD events, utilized an RF ensemble approach with the top 14 selected features. The model demonstrated strong predictive performance, achieving an AUC of 0.944, indicating excellent discrimination between the different CVD events. Despite the limited sample size, the confusion matrix demonstrated good classification performance, correctly classifying most cases of MI and stroke. This outcome suggests that even with a relatively small number of features, the model effectively captured the critical patterns associated with these specific CVD events.

Conclusion: This study demonstrates that the developed models offer robust and accurate 3-year predictions specifically tailored to the Malaysian population. The first model provides a reliable approach to predicting CVD risk within a 3-year timeframe, while the second model effectively classifies specific cardiovascular events over the same period. The results indicate significant improvements over traditional methods, underscoring the potential for these models to enhance CVD risk assessment and intervention in Malaysia.