Heterogeneity in Multimodal AI Feature Engineering Utility for Explaining Change in Dementia

The present work builds upon progress in foundation artificial intelligence (AI) models for thepurpose of supporting the Alzheimer’s Disease Neuroimaging Initiative (ADNI). Open-weight largelanguage models (LLMs) are employed (collaborative inference from eight labs) for AI magneticresonance imaging (MRI) insight extraction and compared in terms of relative feature importance(between-model, between-feature) at the AI-engineered parameter-level (along with other engineeredmeta-features) for relative explanatory utility with regard to participant change in an assessment of mildcognitive impairment (MCI) and dementia. A gradient boosting machine (GBM) suggests severalneuroimage-inferred dimensions extracted by (2 of 8) LLMs outperform all others in explanatory utility(except for baseline clinical dementia rating) including inferred representations of Alzheimer's (AD)probability, small vessel disease burden, cortical atrophy, and cognitive impairment severity. Ak-means-derived cluster feature corroborates among the high importance neuroimage features,suggesting multimodal superiority in the context of the present work. Moreover, given the estimatedvariation in feature importance (beyond the vision superiority evidence) by LLM, input features, andextracted dimensions, the secondary finding suggests heterogeneity with regard to explanatory utilityacross a class of clinically relevant composite dimensions extracted and evaluated between LLM andbetween inputs in the context of a predictive model. Furthermore, unambiguous explanatory importanceis suggested with regard to AD probability and (neuroimage) novelty/atypicality given the estimatedvalue for compression of these AI-inferred dimensions - regardless of source of inference.