Biases in computable phenotypes for identifying patients with type 2 diabetes
- Journal
- BMC medical research methodology (Q1)
- Published
- 20 August 2026
- Study design
- Unclassified
- Evidence level
- Level 5, Expert Opinion (CEBM 5)
- Authors
- Kylie K Harrall, Keith E Muller, William T Donahoo, Olveen Carrasquillo, Juan Pablo Perdomo Rodriguez, Qian Li, et al.
- PMID
- 43035657
- DOI
- 10.1186/s12874-026-02965-3
Why clinicians should know about it
- Picked for Health Informatics (top studies of the week, 22 November 2026): Evaluates demographic bias in EHR computable phenotypes
Abstract
BACKGROUND: We leveraged real-world data from the OneFlorida+ Data Trust to examine sociodemographic biases associated with computable phenotypes (CPs) for type 2 diabetes (T2D) that incorporate prescription and laboratory information. Medical record review was used as a gold-standard to evaluate the discriminative ability of each CP. METHODS: CPs identified patients ≥ 18 years from OneFlorida+ electronic health records (9/1/2018 - 2/29/2020). To clarify the relationships among the partially overlapping CP definitions, we defined four mutually exclusive (ME) groups of individuals: 1) T2D diagnosis only (ME_Claims); 2) T2D diagnosis and antidiabetic prescription (ME_Meds); 3) T2D diagnosis and HbA1c ≥ 6.5% (ME_Labs); and 4) Antidiabetic prescription and HbA1c ≥ 6.5% (ME_Meds_Labs). A split sampling approach and multinomial models were used to examine whether sociodemographic factors were associated with ME category membership. In a substudy, physicians reviewed medical records to verify T2D status, and the discriminative ability of each CP definition was assessed. RESULTS: In urban locations, compared to NH-White persons, Hispanic persons had lower odds of identification when CPs incorporated prescription or laboratory information relative to the ME_Claims definition [OR (95% CI): ME_Meds: 0.50 (0.46, 0.53), ME_Labs: 0.85 (0.79, 0.93), ME_Meds_Labs: 0.40 (0.36, 0.44)]. In rural locations, compared to NH-White persons, there were lower odds of identification by CPs incorporating prescription or laboratory information relative to the ME_Claims definition among NH-Black persons [OR (95% CI): ME_Meds: 0.72 (0.64, 0.82), ME_Labs: 0.52 (0.40, 0.68), and ME_Meds_Labs: 0.55 (0.46, 0.65)] and Other persons [OR (95% CI): ME_Meds: 0.59 (0.45, 0.77), ME_Labs: 0.48 (0.26, 0.87), and ME_Meds_Labs: 0.45 (0.31, 0.67)]. The medical record substudy revealed that the CP requiring a T2D diagnosis and laboratory results had the highest discriminative ability (adjusted AUC = 0.818) and specificity (1.000), while the CP requiring only a T2D diagnosis had the highest sensitivity (0.896). CONCLUSIONS: CPs have varying discriminative ability and may introduce measurement bias by excluding patients with less access to care. The interaction between race-ethnicity and rurality illustrates how CPs using prescriptions and laboratory information may omit certain populations. Differences in sensitivity and specificity highlight the varying utility of CPs across research contexts, such as surveillance studies versus cohort discovery.
Abstract as published, via PubMed.
For healthcare professionals. The summary is generated by AI from the published abstract, and the evidence level is assigned automatically from the study design on the Oxford CEBM hierarchy. Neither is medical advice. Read the full paper before changing practice.