Optimizing decision thresholds and model selection for imbalanced clinical risk prediction: an ensemble framework for single-disease screening

Chronic diseases such as cardiovascular disease, stroke, and cirrhosis impose substantial clinical and economic burdens, yet predictive models often falter in the presence of severe class imbalance and suboptimal decision thresholds. This study proposes an imbalance-ratio–aware, threshold-optimized ensemble for single-disease risk prediction. The framework adapts its learning route to the observed imbalance ratio, integrates synthetic oversampling and cost-sensitive learners when imbalance is pronounced, and fuses calibrated probabilities from the top performers using validation F1-based weights. Decision thresholds are tuned on precision–recall characteristics rather than fixed at 0.5, separating probability estimation from policy and aligning operating points with clinical priorities. For stroke (extreme skew), the ensemble increased recall to 0.6000 and improved F1 to 0.3659, raising G-Mean and MCC to 0.7404 and 0.3519, with an optimized threshold of 0.6880. For heart disease, the model sustained F1 at 0.9065 and lifted ROC-AUC to 0.9328 at a threshold of 0.4339. For cirrhosis, F1 reached 0.8571 with recall 0.8438 and threshold 0.3439. Averaged across tasks, F1, ROC-AUC, G-Mean, and MCC were 0.7098, 0.8956, 0.8348, and 0.6353, respectively. These findings indicate that coupling imbalance-aware training with threshold optimization yields robust, clinically meaningful performance. Future work will extend external validation, utility-constrained thresholding, and domain adaptation to strengthen deployment in real-world screening pipelines.