1110-P: Algorithmic Identification of Atypical Diabetes in Electronic Health Record (EHR) Systems

Abstract

Understanding atypical forms of diabetes may lead to personalized treatment regimens and discovery of novel pathophysiologic mechanisms of diabetes. We aimed to create a high-throughput method for identifying patients with atypical diabetes in large electronic health record (EHR) databases and validate this method in the Mass General Brigham (MGB) Biobank. Patients with likely type 2 diabetes (T2D) were identified using a validated machine-learning (ML) algorithm. “Typical” T2D was filtered out through a “base algorithm” that excluded individuals with body mass index (BMI) ever >30 kg/m2, HDL values ever < 50 mg/dL, and triglycerides ever >150 mg/dL. To remove people with typical type 1 diabetes (T1D) , we tested six additional “branch algorithms,” relying on clinical characteristics, including autoantibodies, medication usage, and T1D diagnosis determined by ML algorithm, resulting in six overlapping cohorts. Charts were then manually reviewed, and diabetes type was classified by two endocrinologists into one of three categories: atypical, not atypical, and indeterminate due to missing information. We identified 1potentially atypical cases, of whom 16 individuals were confirmed to have atypical diabetes after expert review, across the six branch algorithms. The branch algorithm which excluded T1D by removing patients who had ever used outpatient insulin had the highest percentage yield (13 of 27; 48.2%) of atypical diabetes . The 16 atypical cases had significantly lower BMI and higher HDL compared to an unselected group of individuals with T1D or T2D diagnosis by ML algorithm. Compared to the ML T1D group, the atypical group had a significantly higher T2D polygenic score and lower hemoglobin A1c. In summary, we designed an algorithm to identify individuals with atypical diabetes within an EHR database with up to 48% yield which may shed light on the heterogeneity of T2D and help generate cohorts of atypical cases for studies, such as the Rare and Atypical Diabetes Network (RADIANT) . Disclosure V.Chen: Other Relationship; Draeger. S.Cromer: Other Relationship; Johnson & Johnson Medical Devices Companies, Research Support; National Institute of Diabetes and Digestive and Kidney Diseases. C.Han: None. W.G.Marshall: None. S.Emongo: None. T.Majarian: Employee; Vertex Pharmaceuticals Incorporated. J.C.Florez: Consultant; AstraZeneca, Goldfinch Bio, Inc., Other Relationship; AstraZeneca, Merck & Co., Inc., Novo Nordisk. J.M.Mercader: None. M.Udler: None.