Optimizing Alzheimer’s Disease diagnosis with MRI and clinical data via lightweight transformers

Alzheimer’s Disease diagnosis requires integrating neuroimaging and clinical assessments. This study describes a multimodal framework for subject-level AD classification using sagittal MRI and clinical data from the OASIS-1 and OASIS-2 datasets. The framework uses MobileViT as a fixed feature extractor for MRI slices and a ClinicalMLPEncoder for clinical data. Subject-level SMOTE addresses class imbalance. The Clinical Dementia Rating (CDR) is used exclusively for labeling. Four fusion strategies (Intermediate, Cross-Attention, Early, and Late) are compared under subject-level 5-fold cross-validation with regularization, Focal Loss, and Automatic Mixed Precision. Cross-Attention fusion yields AUC of 0.941 on OASIS-1 and 0.903 on OASIS-2 for binary classification, with accuracies of 87.4% and 82.0%. ClinicalMLPEncoder outperforms TabNet. Error analysis indicates that misclassifications occur at the mildest disease stage (CDR = 0.5). Multi-class performance (3-class) produces AUCs of 0.907 and 0.833 on OASIS-1 and OASIS-2. MC-Dropout yields Brier scores of 0.135–0.145. External validation on the independent MIRIAD cohort yielded an AUC of 0.784 ± 0.038, with the standard deviation reflecting variability across five paired fold-ensembles. The Cross-Attention framework with ClinicalMLPEncoder provides a benchmark for multimodal AD diagnosis. Subject-level evaluation and CDR-as-label prevent leakage and provide performance estimates. Future work includes validation on ADNI and longitudinal analysis.