1368-P: A Prediction Model for Prediabetes and Diabetes Using Easily Obtainable Clinical Data

Introduction and Objective: Chronic diseases such as cirrhosis can reduce the accuracy of the hemoglobin A1c as a diagnostic test for diabetes (DM). We aimed to determine if easily obtained clinical data could be used to improve the diagnosis of DM beyond A1c in the general population. Methods: We analyzed 13,800 subjects from NHANES from 2005-2016 who had an A1c and OGTT. Including standard labs and vital signs features with <12% missing data, we split the subjects and applied the machine learning (ML) approach XG Boost to identify predictive features of OGTT 2-hour glucose (2hG) levels ≥140 mg/dL (pre-DM) and ≥200 mg/dL (DM). Results: The rate of DM by A1c and OGTT was 5.3%, by A1c alone was 0.4%, and by OGTT alone was 3.5%. Of those with pre-DM by A1c, 11.8% had 2hG ≥ 200 mg/dL; 1.5% of those without pre-DM had 2hG ≥ 200 mg/dL (A). The most important variables were included in the model: age, height, arm and waist circumference, pulse, blood pressure, fasting glucose, insulin, iron, and triglycerides, A1c, cholesterol, platelets, GGT, creatinine, neutrophil percentage, urine albumin and creatinine, and Poverty Ratio. The AUC of the model vs. the A1c for pre-DM was 0.76 vs. 0.67 and for DM was 0.92 vs. 0.87, respectively (B). For A1c < 6.3% the model had a higher average positive predictive value (boxplots) than the A1c (blue lines) (C). Conclusion: Incorporation of easily obtainable clinical data into a ML model can improve diagnosis of pre-DM and DM. A. Hutchison: None. C. Thomas: None. E. Tasali: None. M.E. Rinella: Consultant; 89bio, Inc, Akero Therapeutics, Inc, Boehringer-Ingelheim, Eli Lilly and Company, Cytodyn, Inventiva Pharma, Echosens, Novo Nordisk, Madrigal Pharmaceuticals, Inc, Intercept Pharmaceuticals, Inc, Sagimet Biosciences. R.G. Mirmira: None. W.F. Parker: None.