Hybrid CA-SAE-AFB and ensemble learning framework for accurate heart disease prediction

Heart disease prediction from structured clinical records requires models that can represent nonlinear feature interactions while avoiding optimistic estimates caused by data leakage and model-selection bias. This study presents a hybrid CA-SAE-AFB framework that combines a sparse autoencoder, adaptive latent-feature re-weighting, logistic regression, XGBoost, out-of-fold stacking, validation-only F1 threshold optimization, and a two-stage decision rule. The study includes a 50-seed stratified analysis for stability assessment and uses leakage-controlled nested stratified 5-fold outer cross-validation with a 3-fold inner loop as the primary evaluation protocol. All preprocessing, representation learning, threshold selection, and meta-learner fitting are restricted to the appropriate training/inner-validation partitions, while outer test folds remain untouched until final evaluation. Under the nested evaluation, CA-SAE-AFB achieved a mean accuracy of 0.8617 (95% CI: 0.8365–0.8869), F1-score of 0.8772 (95% CI: 0.8591–0.8952), MCC of 0.7208 (95% CI: 0.6695–0.7721), and ROC-AUC of 0.9153 (95% CI: 0.8944–0.9361). The repeated-split and nested-CV estimates were closely aligned. Component ablation showed the clearest directional contribution from adaptive feature boosting, whereas the effects of the sparse autoencoder and meta-learner were metric-dependent. On an NVIDIA Tesla T4 GPU, the complete single-sample inference pipeline required 83.18 ms on average over 200 timed runs after 50 warm-up runs. These results support computational feasibility and methodological robustness for further investigation in clinical decision-support settings, while external prospective clinical validation remains necessary.