1305-P: Detection of Diabetes Status and Type in Youth Using EMRs

Diabetes surveillance often requires manual chart review to confirm status and type. This project aimed to derive and validate an EMR-based algorithm using structured data in the era of ICD-10 codes for improving screening efficiency. Youth (<20 years) with potential evidence of diabetes based on in- and out-patient visits were identified from EMRs at 3 Children’s Hospitals participating in the SEARCH Study. Charts were reviewed to determine presence and type of diabetes based on provider notes. Two methods were compared: a multinomial regression evaluating 29 predictors, and a deterministic approach based on ratios of ICD-10 codes (≥2 diabetes codes with the most frequently occurring type-specific codes to indicate type). To evaluate performance, data from two hospitals were used for model building and the third for testing. From 8,752 potential cases, 5,308 true cases were identified; T1D (89.2%), T2D (7.5%), Other types (3.3%). Age, race/ethnicity, ICD-10 codes, and medications were the strongest predictors in the multinomial regression. Performance metrics are shown. Both methods performed well for diabetes diagnosis and for classification of T1D (all metrics ≥0.95). For T2D, the regression model had substantially lower sensitivity (0.59 vs. 0.90), but higher positive predictive value (0.80 vs. 0.62). The ICD-10 ratio model is easier to implement and could be combined with targeted chart reviews to increase efficiency for surveillance efforts.

View largeDownload slide

View largeDownload slide

Close modal