LLM Embeddings Complement Theory-Driven Language Features and Cognitive Scores for APOE4 and Aβ Prediction

LLM embedding models achieved the highest F1 in five of eight tasks for predicting APOE4 carrier status and cerebral amyloid-beta burden in Hebrew- and English-speaking cohorts (IRAP N=134; WRAP N=1,541).

• Why it matters: Preclinical Alzheimer’s disease needs scalable, noninvasive biomarkers, and it remains unclear how modern language representations complement cognitive scores and theory-driven language features for genetic and amyloid prediction.
• What they did: Using Cookie Theft picture descriptions, the study compared pretrained LLM embeddings, standard cognitive test scores, and theory-driven linguistic features, with amyloid-beta burden split by lowest-, median-, and highest-quartile cutoffs.
• The result: An any-positive ensemble outperformed every constituent model and a prior-informed random baseline across all eight tasks, and beat all constituent models in 21 of 24 regional amyloid-beta tasks, with disagreement indicating complementary signals.

AI summary of the abstract · may contain errors · not medical advice · read the paper.