Beyond Accuracy: Reliability-Aware Machine Learning for Handwriting-Based Alzheimer’s Disease Detection

Reliable clinical decision support systems require not only high predictive accuracy but also trustworthy probability estimates and robust uncertainty quantification. However, most medical artificial intelligence (AI) studies primarily emphasize discrimination performance while overlooking systematic reliability evaluation. This study proposes a reliability-aware evaluation framework for Alzheimer’s disease detection that integrates discrimination analysis, statistical validation, probability calibration, uncertainty quantification, robustness assessment, and clinical decision analysis within a unified pipeline. Multiple machine learning classifiers and ensemble configurations were evaluated using repeated stratified cross-validation and assessed through discrimination and calibration metrics. Support Vector Machine achieved the highest ROC-AUC (0.955 ± 0.046), while Extra Trees obtained the highest Accuracy (0.878) and F1-score (0.887). Friedman analysis confirmed statistically significant differences among classifiers (p<0.001). Platt scaling consistently improved probabilistic reliability, whereas Beta calibration demonstrated stable performance under noise, feature perturbation, and reduced-data scenarios. Uncertainty-aware selective prediction increased high-confidence diagnostic accuracy by up to 8.1%, and decision curve analysis demonstrated improved clinical utility. The reliability analysis identified calibration-aware stacking as the most reliable ensemble configuration. An independent cross-dataset evaluation on a heterogeneous Alzheimer’s disease clinical dataset with a substantially different feature space yielded stable discrimination (ROC-AUC = 0.858 ± 0.025) and calibration (ECE = 0.132 ± 0.019) after the STACK_CAL architecture was independently retrained from scratch. These findings provide evidence of the cross-dataset applicability of the proposed reliability-aware strategy across different clinical data modalities, while further prospective and independent validation remains necessary before real-world clinical deployment.