Systematic Modality Ablation of Multimodal Machine Learning for Predicting 24-Month Progression from Mild Cognitive Impairment to Alzheimer's Disease

Multimodal biomarkers have transformed Alzheimer's disease research, but the incremental contribution of individual modalities to predicting progression from mild cognitive impairment (MCI) remains unclear. We systematically evaluated the contribution of demographic, cognitive, genetic, structural imaging, cerebrospinal fluid (CSF), and positron emission tomography (PET) biomarkers using a comprehensive ablation framework. We analyzed 2,430 participants with MCI from the Alzheimer's Disease Neuroimaging Initiative with known 24-month progression status. XGBoost models were trained using combinations of demographic variables, cognitive assessments, apolipoprotein E (APOE) genotype, structural MRI, CSF biomarkers, and PET biomarkers. Performance was evaluated using repeated stratified 5X10 cross-validation, with out-of-fold AUC comparisons and Holm-Bonferroni correction. Sensitivity analyses assessed the effects of missing-data handling. The full multimodal model achieved the highest discrimination (AUC=0.934). Excluding cognitive assessments produced the largest reduction in performance (AUC=0.883, P<0.001). Removing APOE, CSF, or MRI produced only modest reductions (AUC=0.933, 0.931, and 0.932, respectively). PET produced a similarly small reduction in the primary analysis (AUC=0.932), although complete-case analysis indicated that imputation significantly inflated its performance (P=0.005), suggesting that its contribution may be underestimated or obscured by missingness. The baseline clinical model performed near chance (AUC=0.556). These findings establish an evidence-based hierarchy of biomarker contributions and provide a quantitative framework for prioritizing biomarker acquisition and designing cost-effective multimodal prediction models.