Large Language Model-Based Identification of Subjective Cognitive Decline in Electronic Health Records Among Older Adults With Normal Cognitive Testing: Retrospective Cohort Study.
Key claim: A 2-stage LLM pipeline accurately identified documented cognitive concerns consistent with SCD in 13.8% of older adults with normal cognitive testing, with documentation uncommon and selectively associated with clinical and sociodemographic factors.
Abstract
BACKGROUND: Subjective cognitive decline (SCD) typically refers to self- or informant-reported decline in cognition despite the absence of objective impairment on standardized testing. Older adults with documented normal cognitive test performance provide a pragmatic anchor cohort for electronic health record (EHR)-based SCD phenotyping. However, cognitive concerns are primarily recorded in unstructured notes and are inconsistently documented, making it unclear how often and in whom concerns are captured in routine care. OBJECTIVE: This study aims to operationalize EHR-based SCD phenotyping in an objectively normal-testing cohort using a large language model (LLM)-based natural language processing approach and to examine the frequency and clinical and sociodemographic correlates of cognitive concern documentation. METHODS: We conducted an EHR-based observational study of patients aged 65 years or older with a first normal cognitive test recorded in EHR flowsheets between January 2019 and April 2024 in a large health care system. We developed and iteratively refined a 2-stage LLM-based natural language processing pipeline to identify documented cognitive concerns in unstructured notes during the 12 months prior to the index date, and evaluated performance against manual review. We quantified the frequency of documented concerns within this normal-testing cohort and used multivariable logistic regression to assess associations with sociodemographic factors (age, marital status, insurance, and neighborhood Area Deprivation Index), sequentially adjusting for clinical comorbidities and relevant medications. RESULTS: Among 15,750 older adults with normal cognitive test scores, 13.8% (n=2175) had at least 1 documented cognitive concern in the prior year, captured in 1.2% (7394/605,177) of notes. On manual validation, the 2-stage pipeline (Med42-v2-8B screening followed by GPT-4o confirmation) achieved a sensitivity of 0.957, positive predictive value of 0.935, specificity of 0.985, and-score of 0.945 for cognitive concern identification. Documentation was more likely in older individuals, those with commercial (vs Medicare) insurance, and those with neurological and psychiatric conditions. In fully adjusted models, Parkinson disease (adjusted odds ratio [aOR] 5.29, 95% CI 3.60-7.77), traumatic brain injury (aOR 4.63, 95% CI 3.15-6.81), stroke or transient ischemic attack (aOR 3.46, 95% CI 3.15-4.18), epilepsy (aOR 2.52, 95% CI 1.83-3.49), depression (aOR 1.54, 95% CI 1.36-1.75), and excessive alcohol use (aOR 1.48, 95% CI 1.06-2.06) were among the strongest correlates of cognitive concern documentation. In contrast, obesity (aOR 0.73, 95% CI 0.65-0.82), hyperlipidemia (aOR 0.59, 95% CI 0.51-0.67), and residence in more deprived neighborhoods (higher Area Deprivation Index) were associated with lower odds of documented cognitive concerns. CONCLUSIONS: A 2-stage LLM pipeline enabled accurate identification of documented cognitive concerns consistent with SCD among older adults with normal cognitive testing. Documentation was uncommon and selectively captured by clinical and sociodemographic factors, with implications for equity and the validity of EHR-based phenotypes.
Why this matters
LLM-based EHR phenotyping is becoming a core infrastructure layer for longevity and cognitive-aging research, and this study shows a validated pipeline can surface subjective cognitive concerns that standard testing misses in a sizeable minority of older adults. The selective association with clinical and sociodemographic factors signals that both the underlying documentation and any downstream AI tool inherit equity gaps — a measurement-validity issue the trend coverage of AI-in-longevity routinely glosses over.