A Verifiable Agentic Framework for Sequential Multimodal Prediction of Mild Cognitive Impairment Progression

Abstract

Multimodal models can predict progression from mild cognitive impairment (MCI) to Alzheimer’s disease (AD), but most evaluations assume that every modality is immediately available. We developed a retrospective framework using only the Alzheimer’s Disease Neuroimaging Initiative (ADNI) that separates statistical risk estimation, sequential evidence acquisition, language-model explanation, and deterministic verification. Participants with baseline MCI were assigned state-specific landmarks using information visible within a +/−90-day window. The primary simulation cohort included 425 participants with clinical/cognitive, magnetic resonance imaging (MRI)-derived, and cerebrospinal fluid (CSF) evidence; 153 converted to AD within 36 months after the landmark. Calibrated elastic-net models were evaluated with 10 repeated nested five-fold cross-validation. In the common-landmark cohort, adding MRI-derived and CSF features improved the Brier score from 0.158 (95% confidence interval (CI) 0.141–0.175) to 0.138 (0.122–0.155) and the area under the receiver-operating characteristic curve (AUROC) from 0.837 (0.802–0.871) to 0.878 (0.847–0.908). A sequential policy used cross-fitted estimates of individual Brier-score gain to decide whether to request MRI or CSF. The fixed-threshold exploratory analysis requested 40.6% fewer simulated tests than an all-tests policy while meeting the prespecified Brier and AUROC noninferiority margins. Under fully nested threshold selection, the policy requested 46.9% fewer tests and met the AUROC criterion, while narrowly missing the Brier criterion (difference 0.00546, 95% CI 0.00090–0.01003; margin +0.01). In a 40-card-per-model engineering audit, local Qwen3-4B and Qwen3-8B models copied probabilities and cited supplied evidence correctly in 40/40 cards but matched the deterministic action in 16/40 and 15/40 cards, respectively. Execution authority therefore remained with the deterministic policy. Here, agentic refers to the orchestration of modular evidence states, policy checks, verification, and explanation; it does not imply autonomous language-model control. As a retrospective simulation, the framework does not establish prospective clinical utility, real-world cost savings, or external computational reproducibility.