Confidence-Gated Modality Acquisition for Cost-Aware Multimodal Alzheimer's Disease Screening

Multimodal machine learning pipelines for Alzheimer’s disease (AD) screening typically combine structural MRI with routinely collected clinical and cognitive variables, assuming that acquiring and fusing both modalities is worth the added cost. In primary-care and resource-constrained settings, however, MRI is expensive, slow to obtain and often inaccessible, whereas clinical variables are collected at nearly every visit. This paper asks when acquiring MRI is actually necessary, given an automated ensemble of nine tabular classifiers trained on seven routine clinical features (AUROC 0.836, balanced accuracy 0.770 on an external test cohort). We propose a confidence-gated acquisition policy that requests imaging only when the ensemble’s prediction confidence falls below a tunable threshold τ, and otherwise falls back to a fused tabular-imaging prediction. MRI can be avoided for roughly 90% of subjects at τ = 0.55 with an AUROC change of only 0.0002, and the policy degrades gracefully under simulated missing-modality conditions. SHAP analysis identifies the Mini-Mental State Examination (MMSE) score as the most influential predictor.