An Interpretable Multimodal Machine-Learning Model for Predicting Cognitive Impairment from Routine MRI: A Study with External Validation

Existing MRI-based prediction models for cognitive impairment often rely on a single biomarker or simply compare algorithms, act as black boxes, and lack external validation, which limits their clinical translation. We developed and externally validated an interpretable multimodal machine-learning model that classifies a MoCA-defined screening outcome as a binary task: cognitive impairment (CI, MoCA < 26) versus no cognitive impairment (NCI, MoCA ≥ 26). Sixteen candidate features were considered, including AI-derived medial temporal atrophy (MTA), regional white matter hyperintensity (WMH) volumes and proportions, and Fazekas ratings from routine T1WI and T2-FLAIR, together with age and sex. In a single-center cohort of 433 subjects (302 CI, 131 NCI) with a 7:3 training/hold-out split, five algorithms were compared and Random Forest performed best; SHAP-guided recursive feature elimination within five-fold cross-validation then reduced the set from 16 to 11 features. The optimized 11-feature Random Forest reached a cross-validated AUC of 0.883 ± 0.024 on the training folds and an internal hold-out AUC of 0.966, and 0.860 in an independent ADNI cohort of 139 subjects (100 CI, 39 NCI). SHAP identified the left MTA score as the most influential feature, followed by near-ventricular (juxtaventricular and periventricular) WMH burden and the right MTA score. The left MTA score was most strongly correlated with MoCA (r = −0.481). The model may offer an objective, routine-MRI-based screening aid, pending prospective multicenter validation.