A Verification Study: Comparing Human and Large Language Model Approaches to Extract Alzheimer’s Disease Determinants

As Alzheimer’s disease (AD) research continues to expand rapidly, synthesizing evidence on AD determinants across large scientific databases has become increasingly difficult at scale. This verification study compares human manual extraction with a Large Language Model (LLM) approach for identifying AD determinants from 10 systematic reviews retrieved from the Web of Science database. Two parallel extraction workflows were applied to the same source material. The human researcher identified 81 distinct determinants, whereas the LLM identified 74, with 44 human determinants and 41 LLM determinants overlapping across both approaches. While both methods showed strong alignment on well-established behavioral and lifestyle risk factors, the LLM frequently consolidated clinically distinct variables into broader umbrella categories, a pattern termed lossy consolidation. Although the LLM demonstrated major advantages in efficiency and quantitative evidence retrieval, the findings support hybrid evidence-synthesis workflows in which LLM extraction is combined with targeted human validation to preserve determinant-level clinical specificity.