A Machine Learning Model for Multi-Level Classification of Diabetic Peripheral Neuropathy Using Clinical, Lifestyle, and Familial Factors

Abstract

This study aims to develop a robust Machine Learning (ML) framework for multi-level classification of Diabetic Peripheral Neuropathy (DPN) severity by integrating clinical indicators, lifestyle factors, and familial history in patients with type 2 diabetes. The dataset, collected from the Diabetes Research and Treatment Center, Shahid Sadoughi University of Medical Sciences, Yazd, Iran, underwent comprehensive preprocessing including normalization via MinMaxScaler and class rebalancing using the Synthetic Minority Over-sampling Technique (SMOTE). Several ML algorithms — Random Forest (RF), eXtreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), Support Vector Machine (SVM), Multilayer Perceptron (MLP), and Voting Classifier — were implemented and systematically compared. Among these, the RF model achieved the best performance with an accuracy of 86.1%, demonstrating superior stability and interpretability, closely followed by XGBoost. Feature engineering and the incorporation of clinically meaningful composite indices significantly enhanced model performance by capturing complex relationships among physiological and lifestyle variables. Model evaluation based on accuracy, sensitivity, specificity, F1-score, and ROC-AUC confirmed both predictive reliability and clinical applicability. To further enhance interpretability, SHapley Additive exPlanations (SHAP) analysis was conducted using the XGBoost framework due to its higher compatibility with gradient-based explanation methods. The SHAP results confirmed the consistency of feature importance observed in RF, revealing that lower Mean Reflex values, reduced vibration sensitivity (Tuning Fork Test), and higher Body Mass Index (BMI) were strongly associated with severe neuropathy levels. These findings highlight that combining predictive modeling with explainable Artificial Intelligence (AI) approaches can provide transparent, clinically interpretable insights — paving the way for intelligent, explainable decision-support systems in diabetic care.