Atypical diabetes algorithm design

Starting with the 7,147 individuals in the ML T2D group, we developed a “base algorithm” to filter out individuals with “typical” T2D who ever had evidence of metabolic syndrome, cystic fibrosis-related diabetes (CFRD), or laboratory-demonstrated autoimmune diabetes (Fig 1). Specifically, the algorithm removed individuals with HDL values ever less than 50 mg/dL, triglyceride values ever greater than 150 mg/dL, or BMI values ever greater than 30 kg/m², without requirement for these criteria to be met concurrently, to rule out those with metabolic syndrome or insulin resistance as a likely contributor to diabetes onset. Individuals with fewer than three normal BMI lab values in the EHR were excluded to ensure reliable engagement with the healthcare system over time and thus accuracy of phenotype data. Finally, individuals were excluded if they had ICD codes for cystic fibrosis or ever had any positive T1D autoantibodies, including glutamic acid decarboxylase 65-kilodalton isoform (GAD65), islet antigen 2 (IA2), and zinc transporter 8 (ZnT8).

MGBB: Mass-General Brigham Biobank; T2D: type 2 diabetes; BMI: body mass index; HDL: high-density lipoprotein; T1D: type 1 diabetes; NPV: negative predictive value; PPV: positive predictive value.

While we removed patients with positive autoantibodies in the “base algorithm” to filter out T1D cases who were inappropriately included in the ML T2D starting group, many patients never had antibodies checked or could have tested negative but otherwise fit clinically with typical T1D. We therefore next developed and evaluated six “branch algorithms” as different possible methods to rule out patients with likely T1D (Fig 1), leading to six overlapping cohorts of potentially atypical cases.