We adapted two prior phenotyping algorithms, one from eMERGE developed in 2011 [32] and one from Mass General Brigham (MGB) developed in 2014 [33]. The eMERGE algorithm was a rule-based algorithm originally developed at Northwestern University. Three eMERGE sites conducted chart reviews of cases and controls defined by the algorithm with sample sizes ranging from 50 to 100 cases and 44 to 50 controls, demonstrating 98% positive predictive value (PPV) for cases and 100% negative predictive value (NPV) for controls. The eMERGE algorithm was modified to remove the exclusion of charts with at least one ICD9 code for type 1 diabetes to improve sensitivity, to add ICD10 codes, and to add new diabetes medications released since 2011 (referred to as the “modified eMERGE algorithm”). The MGB algorithm was a machine learning-based algorithm using the PheCAP method [33] that had 90% PPV among a chart review dataset that was screen positive for one of 19 phenotypes (prevalence of definite or possible T2D 16%). This algorithm was modified to add ICD10 codes and to add new diabetes medications released since 2014 (referred to as the “modified MGB algorithm”).
