Revisiting Logistic Regression for Diabetes Prediction: An Interpretable and High-Performance Analysis Using a Public Health Dataset

Logistic regression remains one of the most widely applied statistical techniques for modelling binary health outcomes due to its simplicity, interpretability, and robust theoretical foundation. Despite the growing popularity of complex machine learning algorithms in disease prediction, the continued relevance of traditional statistical models warrants empirical reassessment. This study re-examines the effectiveness of logistic regression for predicting diabetes status using a publicly available Kaggle dataset comprising demographic and symptomatic variables. A quantitative, cross-sectional research design was adopted, with diabetes status coded as a binary outcome. Model assumptions were assessed through multicollinearity diagnostics, while overall performance was evaluated using deviance statistics, Akaike Information Criterion (AIC), pseudo-R², and classification measures derived from a confusion matrix. The fitted logistic regression model demonstrated excellent predictive performance, achieving an overall accuracy of 93.3%, sensitivity of 94.4%, and specificity of 91.5%. Multicollinearity was not found to be a concern, supporting the stability of coefficient estimates. Several predictors, including polyuria, polydipsia, polyphagia, genital thrush, irritability, partial paresis, age, and gender, emerged as statistically significant determinants of diabetes status. The estimated odds ratios confirmed the strong clinical relevance of key symptomatic indicators, particularly polyuria and polydipsia. The findings highlight that logistic regression can rival more complex machine learning models in predictive accuracy while maintaining superior interpretability and lower computational demands. This study reinforces the continued value of logistic regression as a reliable and transparent tool for diabetes risk prediction, clinical decision support, and public health screening initiatives.