Development and validation of a multivariable Prediction Model for Pre-diabetes and Diabetes using Easily Obtainable Clinical Data

ImportanceThe hemoglobin A1c and fasting plasma glucose (FPG) have known limitations for diabetes diagnosis, but models to identify individuals who would benefit from 2-hour oral glucose tolerance testing (OGTT) are limited. ObjectiveTo determine if OGTT-only diagnosed diabetes has comparable outcomes to A1c- or FPG-diagnosed diabetes and if standard clinical features could be leveraged to identify undiagnosed diabetes. DesignMultivariable prediction model development and validation. SettingUS National Health and Nutrition Examination Survey (NHANES) data with corresponding US National Center for Health Statistics (NCHS) mortality data. ParticipantsOf 105,862 NHANES subjects from 1999 to 2016, we identified 13,800 subjects with FPG, A1c, or OGTT results (11,550 with mortality data) and 92,062 other subjects (53,255 with mortality data). ExposureOGTT-diagnosed diabetes Main Outcomes and MeasuresThe primary outcomes were association of mortality with diabetes diagnostic approach and models to diagnose diabetes. We used a gradient boosted machine decision tree to predict diabetes from standard clinical features. In the test set, we compared the AUROC and the net benefit by decision curve analysis to A1c, FPG, and a combination of the two. We performed survival analysis based on method of diabetes diagnosis and diabetes model predictions. ResultsThe rate of OGTT-only diabetes was 1.34%. Subjects with OGTT-only diabetes had equivalent risk of mortality compared to subjects with FPG- or A1c-diagnosed diabetes after adjusting for age, sex, and race/ethnicity. A model using the A1c and standard clinical features (A1c+ model) outperformed the A1c to exclude diabetes (Sensitivity at Youdens Index: 0.72 vs. 0.37). Adding FPG to that model (A1c/FPG+) outperformed FPG for excluding diabetes (Sensitivity: 0.87 vs. 0.48). Subjects with A1c/FPG+-predicted diabetes but sub-diagnostic A1c and FPG had equivalent mortality (HR=8.2, p<2*10-16) to those with A1c or FPG-diagnosed diabetes (comparison p<0.17). Conclusions and RelevanceDiabetes diagnosed by OGTT alone has equivalent mortality to A1c and FPG-diagnosed diabetes. A model using standard clinical features can bolster the A1c and FPG to identify potentially undiagnosed diabetes. Model predictions associated with mortality equivalently to having a diabetes diagnosis. Implementation of a clinical decision support tool could improve diagnosis of diabetes and lead to earlier interventions. Key PointsO_ST_ABSQuestionC_ST_ABSAre individuals with diabetes only diagnosable by oral glucose tolerance test (OGTT) at similar risk of mortality as those diagnosed by hemoglobin A1c or fasting plasma glucose (FPG)? Can readily-available clinical data improve diagnosis? FindingsStratifying NHANES subjects by OGTT-only (1.34%) vs A1c or FPG diagnosis (4.13%) of diabetes found that OGTT-only diagnosis had equivalent mortality. A model combining A1c and standard clinical features (A1c+) had superior AUROC for diabetes compared to the A1c alone. The addition of FPG (A1c/FPG+) had superior AUROC compared to the FPG and A1c+. The A1c/FPG+ predictions strongly associated with mortality in subjects with sub-diagnostic A1c and FPG. MeaningIncorporation of clinical features can improve diabetes diagnosis missed by A1c and FPG. Model prediction of diabetes associates with mortality.