MM-Fusion: multimodal fusion for Alzheimer's screening using structural MRI and demographic data

Early detection of Alzheimer's disease (AD) from routine MRI remains difficult because many automated methods either ignore clinical context (e.g., demographics) or are too complex for practical use; for this reason, we propose a multimodal fusion model (MM-Fusion), a compact architecture that combines a lightweight deep network for MRI slices with a small multilayer perceptron (MLP) for readily available demographics (age, MMSE, eTIV, nWBV, education etc). The model outputs slice-level probabilities that are aggregated into patient-level screening decisions via mean probability and majority vote. Evaluation is done on OASIS dataset with a clinically aligned binary endpoint (control vs. any dementia) under a strict patient-level split, MM-Fusion achieves per-slice accuracy 0.874 and balanced accuracy 0.804, with strong separation between classes as measured by area under the receiver operating characteristic curve (ROC-AUC = 0.953) and area under the precision-recall curve (PR-AUC = 0.870). At the patient level (mean aggregation), accuracy is 0.868 and balanced accuracy is 0.794, with ROC-AUC 0.952 and PR-AUC 0.870. These results show that a simple MM-Fusion model can provide clear, threshold-tunable screening decisions suitable for clinical workflows while remaining efficient and reproducible.