Identify patient-, informant, or clinician-reported SCC from longitudinal EHR data using a large language model (LLM)-based pipeline.

 

Compare baseline demographic and clinical characteristics between individuals with and without documented SCC.

 

Estimate the association between documented SCC and incident mild cognitive impairment (MCI)/dementia using competing-risk regression models.

 

Methods: 

Study design: Retrospective cohort study using longitudinal EHR data from a large academic healthcare system (Mass General Brigham).

Population: Adults ≥65 years with a first documented normal cognitive test (MMSE, MoCA, Mini-Cog, SLUMS, or AD8) between January 1, 2010, and December 31, 2020. Cognitive test score and date were extracted using GPT-OSS 120B; extraction agreement on 200 manually annotated cases: κ = 0.911 and 0.903, respectively.

Exposure: Patient-, informant-, or clinician-documented SCC within 1 year prior to the index cognitive test, extracted from free-text clinical notes using GPT-OSS 120B (sensitivity 88.9%, PPV 84.2%, F1 86.5%).

Outcome: Incident MCI or dementia at least one year after the index date, identified using ICD diagnosis codes.

Statistical analysis:

 

 

Results: 

Study cohort

Baseline differences

Main finding: Documented SCC were associated with a 3.27-fold higher risk of incident MCI/dementia.

 

 

Conclusion: 

 ✅ Among older adults with normal baseline cognitive testing, documented subjective cognitive concerns were associated with more than a threefold higher risk of subsequent MCI or dementia, independent of demographic and clinical factors.

✅ SCC captured in routine EHR documentation, identified at scale using LLMs, may serve as an early, accessible, and clinically meaningful marker for identifying individuals at elevated risk of cognitive decline.

✅ These findings support further evaluation of scalable NLP-based SCC detection to improve risk stratification and facilitate earlier identification of older adults at increased risk of cognitive decline.