Introduction and Objective: People with mental illness have excess mortality from somatic comorbidities including T2D due to delayed diagnosis and undertreatment. The objective was to develop a machine learning (ML) model to predict the development of T2D in people with mental illness from data routinely collected in the Epic electronic health record (EHR). Methods: On a per-visit basis, all adult patients under the purview of the psychiatric hospital service of the Capital Region of Denmark with an index date (hospital discharge or outpatient visit) between June 30, 2018, and Dec. 31, 2021, were included in the training set. The validation set included patients under the purview of the adjacent Region Zealand. Prevalent cases were excluded (diagnosed T2D, use of antidiabetic medication, or ≥2 diagnostic glucose values). The outcome was T2D within 1, 2, or 3 years after index date based on WHO diagnostic criteria, censoring at death or end of follow-up. Model features included demographic and clinical factors such as ICD-10 diagnoses, lab values, medications, number of admissions, and lifestyle factors captured in the EHR. Two model types were used (1) logistic regression with elastic net regularization and (2) XGBoost. Results: The training set included 2.31 M visits on 62 K patients; the validation set 622 K visits on 24 K patients. The 3-year incidence of T2D was 2.8%. The selected model achieved an AUROC of 0.90 (95% CI 0.88, 0.92). Among those whose calculated risk was in the top 5%, 28% developed T2D and 53% of cases were identified. Among the top 1%, 55% developed T2D and 21% of cases were identified. Thus, patients identified by the model had 10-20 times increased risk of T2D. The most strongly contributing factors were HbA1c, BMI, triglycerides, age, and plasma glucose. Conclusion: The ability of this ML-developed model to identify those at high risk for T2D among people with mental disease will allow targeted testing of established and novel, mental disease-specific T2D prevention modalities in this underserved group. Disclosure S. Snitker: Employee; Ended; Novo Nordisk A/S. Stock/Shareholder; Current; Novo Nordisk A/S. L. Hassel-Pflugh: None. M.E. Benros: None. Funding Lundbeck Foundation (R383-2022-285)
Snitker et al. (Fri,) studied this question.