ENHANCING DIABETES PREDICTION WITH INTERPRETABLE MACHINE LEARNING: A COMPARATIVE ANALYSIS OF ADDITIVE–MULTIPLICATIVE NEURAL NETWORKS AND KOLMOGOROV–ARNOLD NETWORKS

This study investigates the effectiveness of machine learning (ML) models in diagnosing diabetes and identifying the most influential predictors using the PIMA Indians Diabetes dataset. A particular emphasis is placed on novel neural network architectures, especially the Additive and Multiplicative Neurons Network (AMNN), introduced as a key innovation in this work. The dataset underwent comprehensive preprocessing, including handling missing values, feature scaling, and addressing class imbalance via the SMOTE algorithm. To interpret the importance of predictors, five feature selection techniques (Correlation, Boruta, MRMR, RFE, Random Forest) and two explainable AI (XAI) tools (SHAP and LIME) were applied. A total of eight machine learning algorithms were tested and evaluated based on accuracy, recall, F1-score, and AUC-ROC. Among all models, AMNN achieved the best performance, with an accuracy of 0.7576, recall of 0.7576, F1-score of 0.7618, and AUC-ROC of 0.8206. MLP-2 and XGBoost also showed competitive results. Kolmogorov-Arnold Networks (KAN), while not outperforming other models, demonstrated moderate success and offered interpretability advantages due to its flexible activation structure. Consistently, glucose, BMI, age, and pregnancy count were found to be the most significant predictors across feature selection and XAI evaluations. These results align with existing clinical insights into diabetes risk. In conclusion, this study highlights the potential of the AMNN model as a powerful and interpretable tool for early diabetes detection. These findings suggest that AMNN offers a compelling balance between performance and interpretability, making it suitable for real-world medical applications. The integration of feature selection and XAI techniques supports model transparency, paving the way for its application in clinical decision-making. Future work should focus on enhancing generalizability through larger datasets and hybrid modeling strategies.