Logistic Regression Analysis of Diabetes Risk Factors and Prediction

Abstract

: Diabetes is a major and growing global health threat, making early prediction of high-risk individuals crucial. While complex machine learning models have been explored, the interpretability of logistic regression remains highly valuable for clinical insight. This study aimed to identify key predictors of diabetes onset using a logistic regression model based on the PIMA Indians Diabetes Dataset (PIMA IDD). A backward stepwise regression method, guided by the likelihood ratio test, was employed to optimize the model, and the classification threshold was adjusted via the Receiver Operating Characteristic Curve to maximize Youden's index. The results identified plasma glucose level, body mass index (BMI), and the number of pregnancies as the most statistically significant positive predictors (p < 0.001). By adjusting the classification threshold to 0.292, the model's sensitivity significantly improved from 54.2% to 76.6%, substantially reducing the false negative rate. The findings demonstrate that logistic regression provides a robust and interpretable framework for diabetes prediction, effectively balancing model performance with clinical utility. This study highlights the importance of threshold tuning in aligning predictive models with the priority objectives of clinical screening.