Predicting Type 2 Diabetes Based on Multiclass AUC from Multinomial Logistic Regression

• Activity Number: 246 - Improved Disease Classification Through Extensions of ROC Curve Estimation and Biomarker Characterization
• Type: Contributed
• Date/Time: Tuesday, August 9, 2022 : 8:30 AM to 10:20 AM
• Sponsor: Section on Medical Devices and Diagnostics
• Author(s): Paul G Wakim* and Hima Tallam and Daniel C Elton and Sungwon Lee and Perry J Pickhardt and Ronald M Summers
• Companies: NIH Clinical Center and NIH Clinical Center and NIH Clinical Center and NIH Clinical Center and University of Wisconsin School of Medicine and Public Health and NIH Clinical Center
• Keywords: Type 2 diabetes; predictors; CT scan; logistic regression; multiclass AUC; clinical factor

Abstract

The objective of this exploratory, retrospective cohort study on 7112 patients is to identify the optimal set of predictors of Type 2 diabetes from among 5 clinical factors and 23 measures from Computed Tomography (CT) scans of the pancreas, as well as outside the pancreas.

The nominal response variable is a classification from among the following 5 categories of Type 2 diabetes: 1) Nondiabetic 2) Diabetic, with the date of the CT scan 2500 or more days before the diabetes diagnosis 3) Diabetic, with the date of the CT scan within 2499 days before the diabetes diagnosis 4) Diabetic, with the date of the CT scan within 2499 days after the diabetes diagnosis 5) Diabetic, with the date of the CT scan 2500 or more days after the diabetes diagnosis

With dichotomous response variables, one measure of logistic regression model performance is the Area Under the Receiver Operating Characteristic (ROC) Curve, or AUC. This presentation addresses a method that extends the AUC obtained from models on binary outcomes to the case where the response variable involves 5 nominal categories. The Multiclass AUC is based on work by Hand and Till, published in 2001 in Machine Learning.