A Multisource Explainable AI Framework for Cognitive Deviation Assessment in Aging Brains: A Cross-Dataset Generalization Study

Abstract

Deep-learning systems for cognitive assessment often report strong within-dataset performance without establishing whether the learned representation transfers to independent cohorts. This study evaluates the Hybrid Explainable Cognitive Deviation Tracking (HE-CDT) framework as an evidence-auditable research prototype integrating structural MRI staging, questionnaire-derived behavioral risk, model confidence reporting, explainability, and cross-dataset validation. Six MRI configurations—a custom CNN, MobileNetV2, EfficientNetB0, DenseNet121 baseline, optimized DenseNet121, and a soft-voting ensemble—were compared on a four-class development cohort. The optimized DenseNet121 achieved 79.26% accuracy, 0.8282 macro F1, and 0.9350 ROC-AUC on the internal holdout, while the ensemble achieved the highest internal AUC of 0.9477. Without retraining, optimized DenseNet121 achieved 37.84% accuracy, 0.2705 macro F1, and 0.5441 AUC on the independent OASIS cohort, exposing substantial domain shift.
A methodological audit was additionally performed for dataset provenance, split leakage, class-level artifacts, and fusion-score design. Patient/group-aware splitting is documented where identifiers are available, but augmentation-family isolation could not be fully verified from the recovered manifests. Likewise, a standalone confusion matrix and stylized ROC artifact could not be reconciled with the verified aggregate results; they are therefore excluded from final quantitative claims rather than reconstructed. To remove a conceptual bias in the legacy fusion rule, confidence is separated from deviation severity and retained as a reliability indicator, while a probability-weighted MRI severity formulation is proposed for future calibrated fusion. The study demonstrates that transparent negative external-validation findings and evidence-integrity auditing are essential for credible cognitive-imaging AI research.