LLM Representations Rival Traditional Predictors and Enhance Multimodal Prediction of APOE4 and Aβ-Burden
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Spontaneous language offers a scalable, non-invasive approach for identifying biological indicators of preclinical Alzheimer’s disease (AD). We examined prediction of APOE4 carrier status and cerebral amyloid-β (Aβ) burden using Cookie Theft picture-descriptions from two cohorts, including Hebrew- and English-speakers (IRAP, N=134; and WRAP, N=1,541). Aβ burden was classified across lowest-quartile, median, and highest-quartile cutoffs, capturing accumulation below conventional amyloid-positivity thresholds. We compared embeddings from pretrained large language models (LLMs) to standard cognitive test scores, reflecting current clinical assessment, and theory-driven linguistic features, reflecting traditional language analysis. Embedding-based models achieve the highest F1 among constituent models in five of eight tasks for APOE4 status and global Aβ. High prediction disagreement suggests complementary signals between approaches. An any-positive ensemble significantly outperforms every constituent model and a prior-informed random baseline across all eight tasks, and is also significantly superior to all constituent models in 21 of 24 regional Aβ tasks. These findings highlight the complementarity of LLM-based, theory-driven, and cognitive measures for predicting AD-related genetic risk and pathology, and supports development of early screening approaches.