Comparison of logistics regression models in early detection of diabetes risk

Abstract

Diabetes is a global public health problem affecting 463 million people, and recent studies project a 51% increase in the number of affected individuals by 2045. The rate of asymptomatic pre-diabetic individuals is also high, accounting for approximately 84% of cases, hindering timely intervention and treatment before the disease progresses to severe complications or even death. Early diagnosis of the disease proves beneficial in this scenario, and data science can contribute to achieving it. The aim of this study is to propose early prediction models for diabetes using supervised methods of Binary Logistic Regression and Multilevel Binary Logistic Regression, assessing which model yields more accurate results. This study builds upon previous research where various methodologies were applied, but none utilized a multilevel approach. Responses from a questionnaire administered to patients—both diabetic and healthy—at the Sylhet Diabetes Hospital in Bangladesh were utilized in the modeling process, containing inquiries related to symptoms commonly associated with diabetes diagnosis. This study yielded models with satisfactory results, indicating that multilevel modeling outperforms conventional Logistic Regression