Multinomial Logistic Regression for Diabetes Prediction using LASSO and Elastic Net Regularization

Abstract

Diabetes is a chronic metabolic disorder characterized by persistently elevated blood glucose levels and is associated with severe long-term health complications if not detected early. Accurate classification of individuals into Non-Diabetic, Prediabetic, and Diabetic categories are therefore essential for effective prevention and clinical intervention. This study is a multiclass diabetes prediction model by using Multinomial Logistic Regression integrated with LASSO and Elastic Net regularization techniques. A dataset of 1,000 patient records containing demographic and biochemical variables was analysed following data cleaning, normalization, and an 80:20 train–test split. Model training and hyperparameter tuning were conducted using 10-fold cross-validation. Performance was evaluated by using accuracy, Cohen’s Kappa, F1-score, Brier score, macro-AUC, and information criteria. Elastic Net model achieves the highest accuracy of 0.9141, the highest Cohen’s Kappa of 0.6340, the highest macro sensitivity of 0.6607 and the highest macro balanced accuracy of 0.7303 among the models. In addition, it consists of fewest variables and provided the most balanced and stable result of AIC, BIC, Pseudo-R2, F1-Macro, F1-Weighted, Brier score and Macro specificity which is 387.8591. 401.9204, 0.5556, 0.9559, 0.5737, 0.8909 and 0.8665 respectively. The results indicate that Elastic Net regularization provides the most stable and balanced classification performance while effectively handling correlated variables. Key variables that influenced diabetes status include body mass index, age, HbA1c, triglycerides, and cholesterol. The proposed approach offers an interpretable and reliable framework for multiclass diabetes classification and early risk stratification.