pipette
ESEspañol

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

M. Zadok, N. Calderon, Y. Wald, N. Rein, I. Ganmore, O. Lesman-Segev, G. Almog, T. Shamir, R. Ackler, S. Y. Leviner, G. Noimark, G. Levy, K. D. Mueller, R. Ravona Springer, M. Schnaider Beeri, R. Reichart

Preprint

In the authors' words

Spontaneous language offers a scalable, noninvasive approach for identifying biological indicators of preclinical Alzheimer's disease (AD). We examined prediction of APOE4 carrier status and cerebral amyloid-beta burden using Cookie Theft picture descriptions from two cohorts of Hebrew- and English-speaking participants (IRAP, N=134; WRAP, N=1,541). Amyloid-beta burden was classified using lowest-quartile, median, and highest-quartile cutoffs, capturing accumulation below conventional amyloid-positivity thresholds. We compared embeddings from pretrained large language models (LLMs) with standard cognitive test scores, reflecting conventional clinical assessment, and theory-driven linguistic features, reflecting traditional language analysis. Embedding-based models achieved the highest F1 scores among constituent models in five of eight tasks involving APOE4 status and global amyloid-beta burden. High prediction disagreement suggested complementary signals across approaches. An any-positive ensemble significantly outperformed every constituent model and a prior-informed random baseline across all eight tasks, and significantly outperformed all constituent models in 21 of 24 regional amyloid-beta tasks. These findings highlight the complementarity of LLM-based, theory-driven, and cognitive measures for predicting AD-related genetic risk and amyloid pathology, and support further investigation of their potential for early screening approaches.

Main resultLimitation the authors admit

Appeared: Wednesday, November 11. medRxiv. Preprint, not yet peer-reviewed.

DOI: 10.64898/2026.11.09.26367620