Characterizing cognitive concerns in subjective cognitive decline using large language models and electronic health record notes
Taiyang Chen, Justin Wang, Sheril Varghese, Jiazi Tian, Rebecca E. Amariglio, Li Zhou, Gad A. Marshall, Liqin WangAbstract
INTRODUCTION
Subjective cognitive decline (SCD) refers to self‐ or informant‐reported cognitive concerns despite normal objective testing. These concerns are often documented as free text in electronic health records (EHRs), limiting systematic analysis. We characterized cognitive concern domains and assessed differences across demographic subgroups.
METHODS
We studied 2175 individuals aged ≥65 years with an EHR‐derived SCD phenotype. Using large language models with expert review, we derived a taxonomy of concern themes and sub‐themes and assigned themes to note excerpts. We examined theme frequencies and compared distribution across demographic groups.
RESULTS
We identified 13 themes and 40 sub‐themes. Theme‐level classification showed strong performance (micro‐F1 = 84.8%). Memory Issues was the most common theme (68.5%), with concerns spanning multiple cognitive domains. Theme distribution varied by age, sex, marital status, and education.
DISCUSSION
EHR notes captured heterogeneous, multi‐domain cognitive concerns among adults with an EHR‐derived SCD phenotype. This approach enables scalable, clinically interpretable profiling of concerns that may support early risk stratification.