P4‐326: Automated Machine Learning Methods to Dectect Undiagnosed Cognitive Impairment Using Electronic Medical Records
Year: 2016
Type: article
Study design: Observational Study
Abstract: Elderly individuals are at risk for developing dementia and when undiagnosed, significantly increases health-care costs and reduces quality of life. However, screening for cognitive impairment is impractical and time-consuming in the clinical setting. Therefore, we seek to develop an automated method of topic modeling to detect words and phrases—referred to as topics—associated with dementia and thereby identify undiagnosed dementia. We used a Latent Dirichlet Allocation (LDA) topic-modeling algorithm with a stable topic identification method (sLDA) to identify topics (i.e. recurring themes) in a collection of notes without manual annotations. To capture topics related to early dementia, we applied sLDA to clinic and home-visit notes (n=871,236) of 1,861 cases (i.e., ICD-9 dementia diagnoses) and 9,305 controls (i.e., no dementia diagnoses or not on anti-dementia medications (matched [5:1] to cases by age, gender, and comorbidity) in VA EMR. For cases, notes were analyzed from 6 months to 3 years prior to diagnosis and for controls during the same approximate timeframe. Dementia specialists subsequently reviewed the same notes in ten cases and ten controls (blinded), diagnosing dementia using DSM-V. In the period prior toICD-9 dementia diagnosis, sLDA identified many topics in cases that indicated dementia: e.g. a topic containing the terms “dementia,” “cognitive impairment,” and names of anti-dementia medications appeared in notes of 74.6% of cases versus 13.46% of controls (OR=18.88; 95% CI 16.74, 21,28). A total of 877 unique and stable topics were generated with nine topic had OR>3.0. Agreement between ICD-9 dementia diagnosis and clinician-assigned diagnosis was high (Kappa=0.810, 95% C.I. 0.571, 1.0). We used topic modeling to detect topics in clinical notes that are more likely to be associated with dementia cases than controls. Dementia symptoms were documented in notes many months/years prior to ICD-9 dementia diagnosis. In a subset of subjects, we demonstrated strong agreement between previously assigned ICD-9 dementia diagnoses and manual chart reviews by dementia experts. These studies demonstrate the feasibility of using these topics in a separate group of subjects without previous ICD-9/10 dementia diagnoses, to identify individuals who may have a high probability of undiagnosed dementia. These studies will have implications for clinical practice.
Source: Alzheimer s & Dementia
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: bronze
OpenAlex ID: W2537101226