Systematic Modality Ablation of Multimodal Machine Learning for Predicting 24-Month Progression from Mild Cognitive Impairment to Alzheimer’s Disease
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
INTRODUCTION
Multimodal biomarkers have transformed Alzheimer’s disease research, but the incremental contribution of each modality to predicting progression from mild cognitive impairment (MCI) remains unclear. We systematically evaluated individual biomarker modalities using a comprehensive ablation framework.
METHODS
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, cerebrospinal fluid (CSF) biomarkers, and PET biomarkers. Performance was evaluated using repeated stratified 5×10 cross-validation; discrimination was assessed using the area under the receiver operating characteristic curve (AUC), with pairwise comparisons via DeLong’s test on single-pass out-of-fold predictions, Holm-Bonferroni corrected. Complete-case analyses evaluated the impact of missing data and median imputation.
RESULTS
The full multimodal model achieved the highest discrimination (AUC=0.934). Excluding cognitive assessments produced the largest reduction (AUC=0.883, P<0.001 vs. multimodal), the only comparison to remain significant after correction. Removing APOE, CSF, or MRI produced only modest, statistically indistinguishable reductions (AUC=0.933, 0.931, 0.932). Removing PET produced a similarly small reduction in the primary analysis (AUC=0.932), but complete-case sensitivity analysis showed imputation significantly inflated this estimate (P=0.005); PET’s true contribution may exceed the other three. The baseline clinical model performed near chance (AUC=0.556).
DISCUSSION
Cognitive assessment contributes substantially more predictive information for 24-month progression than structural imaging, molecular biomarkers, or genetic risk, whose removal produces only modest, largely statistically indistinguishable performance loss — with the exception of PET, whose apparent equivalence may be an imputation artifact. These findings establish an evidence-based hierarchy of biomarker utility and provide a quantitative framework for designing cost-effective prediction models and prioritizing biomarker acquisition in clinical research and multimodal AI systems.
