DPN poses a major challenge in diabetes management, affecting a substantial proportion of individuals with long-standing diabetes. Although early detection can prevent irreversible nerve damage, current diagnostic approaches often fail to identify the disease during its subclinical or early symptomatic stages [6]. Traditional assessments rely heavily on subjective clinical judgment, which limits reproducibility and scalability across patient populations [7].

ML offers a promising alternative by leveraging complex datasets encompassing clinical, physiological, lifestyle, and genetic dimensions. These algorithms can learn intricate, non-linear relationships and interactions that are beyond the reach of traditional regression-based analyses [8]. By incorporating diverse patient information—including reflex assessments, vibration tests, BMI, educational level, and family history—ML models can provide a more holistic and accurate prediction of DPN severity.

This research, therefore, investigates how ensemble-based ML algorithms (specifically RF and XGBoost) can improve the precision and robustness of DPN classification. The motivation stems from the need to translate data-driven insights into clinically interpretable models that can assist healthcare professionals in making timely, evidence-based decisions. Achieving this goal would mark a significant step toward early diagnosis, personalized intervention, and prevention of diabetic neuropathy and its associated complications.

Accurate early prediction of DPN onset and progression remains a key challenge in diabetic care, as many patients remain asymptomatic until substantial nerve damage occurs. ML approaches have shown great promise in addressing this issue by integrating clinical, sensory, and biometric indicators to identify patients at risk before irreversible damage develops.

Ma et al. [7] conducted a systematic review of ML-based methods for DPN prediction and emphasized that incorporating biometric and clinical measurements—such as reflex tests, vibration assessments, and metabolic markers—significantly enhances model performance. Similarly, Sheikh et al. [11] utilized plantar pressure distribution data in combination with advanced ML algorithms to predict peripheral neuropathy with high accuracy, illustrating the value of integrating physiological signals with clinical data.

Collectively, these studies highlight the capability of ML techniques to detect neuropathic abnormalities earlier than traditional screening methods, enabling timely intervention and improved patient outcomes.

Predicting and preventing DFUs represents one of the most critical applications of ML in DPN management, as DFUs often lead to infection, hospitalization, and amputation. Ensemble and deep learning models have demonstrated particular strength in improving DFU prediction accuracy by processing high-dimensional and heterogeneous data.

Ma et al. [12] evaluated several ML algorithms—such as RF and SVM—for DFU classification and achieved notable improvements in diagnostic accuracy and computational efficiency. Their findings showed that ML-assisted diagnostic systems could reduce diagnostic time, facilitate early intervention, and improve treatment outcomes, particularly in high-risk diabetic populations.

Despite remarkable progress, several methodological and practical challenges persist in applying ML to DPN prediction and classification. A major limitation involves data quality and heterogeneity, especially in developing regions where access to comprehensive, high-quality medical data is limited. Incomplete or noisy data can reduce model reliability and generalizability, particularly in real-world healthcare settings.

Almutairi et al. [13] highlighted that data incompleteness and imbalance remain key barriers to implementing ML-based prediction systems in clinical practice. Moreover, many existing models focus solely on biochemical or demographic variables, overlooking the potential contribution of lifestyle and familial factors—two dimensions that are crucial for personalized prediction. The lack of multi-modal data integration and insufficient model interpretability also restrict their translation into routine clinical workflows.

Recent advances in ML and data science have opened new avenues for developing more comprehensive and interpretable DPN prediction models. Emerging approaches such as transfer learning and federated learning allow model training on decentralized datasets while addressing data privacy and scarcity concerns. These frameworks enhance model generalizability across institutions and patient populations.

Liu et al. [1] developed a hybrid model combining clinical and biometric data, achieving high predictive accuracy while maintaining interpretability. Similarly, multi-modal ML frameworks that integrate structured clinical data, imaging, and lifestyle indicators are gaining traction as promising solutions for personalized neuropathy prediction.

However, despite these advances, there remains a clear need for comprehensive, population-specific models that integrate clinical, lifestyle, and familial data within an interpretable ML framework. This study aims to contribute to this growing body of research by designing and validating such a model for multi-level classification of DPN severity.

This study used electronic medical records of 1,204 patients with type 2 diabetes, collected between 2019– 2022 at the Yazd Diabetes Research and Treatment Center, Shahid Sadoughi University of Medical Sciences, Iran. The initial dataset contained 70 variables encompassing clinical, neurological, metabolic, lifestyle, and familial factors related to DPN.

Data sources included:

I. Clinical features: BMI, blood pressure, and vascular assessments.

II. Neurological tests: achilles reflex, vibration perception (tuning fork), and muscle cramp sensation.

III. Lifestyle indicators: education, occupation, smoking, and alcohol use.

IV. Familial history: diabetes, neuropathy, and cardiovascular disease in first-degree relatives.

V. Neuropathy assessment: michigan Neuropathy Screening Instrument (MNSI), integrating symptom-based and physical examination scores.

The MNSI score was used as the target variable, categorized into four severity levels: 1) 0: No neuropathy, 2) 1: Mild, 3) 2: Moderate, and 4) 3: Severe.

To ensure data reliability, the following preprocessing steps were applied:

I. Missing data handling: records with >20% missing values were excluded; remaining gaps were imputed (mean for numerical, mode for categorical).

II. Outlier removal: the Interquartile Range (IQR) method was applied.

III. Normalization: continuous features were scaled to the [0,1] range using MinMaxScaler.

IV. Class balancing: class imbalance was corrected using Synthetic Minority Over-sampling Technique (SMOTE) to ensure proportional representation across severity levels.

After preprocessing, a clean dataset of 1,204 complete samples was obtained for modeling.

Feature selection followed a structured three-phase process:

I. Initial reduction: variables with >80% missing data were removed, reducing features from 70 to 20.

II. Feature importance analysis: ensemble algorithms (RF, XGBoost) were used to evaluate feature relevance via Gini impurity and information gain.

III. Final selection: based on cross-validation and clinical significance, 8 features were retained:

− Gender − Education level − Family history (diabetes/neuropathy) − BMI − Muscle cramp sensation − Tuning fork test − Left and Right Achilles reflexes

Feature engineering was conducted to extract higher-level and interaction-based indicators. All potential composite and interaction features were generated and statistically evaluated. The following engineered features were retained:

I. Metabolic Risk Score (MRS): combines BMI, gender, and family history.

II. Reflex Symmetry Index (RSI): difference between left and right Achilles reflexes.

III. Sensory Composite Index (SCI): mean of tuning fork and reflex scores.

IV. MNSI composite score: integrated reflex and symptom metrics.

These engineered variables improved both model interpretability and predictive consistency.

Six algorithms were trained for comparison: RF, XGBoost, LightGBM, SVM, MLP, and a Voting Classifier (ensemble). Each model was trained using 5-fold cross-validation, with hyperparameters optimized through Grid Search CV.

Models were assessed using multiple performance metrics: accuracy, sensitivity (Recall), specificity, F1-Score, and AUC-ROC. This evaluation ensured robust performance measurement across all neuropathy severity levels without overemphasizing specific model outcomes.

To enhance the interpretability of the predictive framework, two complementary techniques were applied. First, a Feature Importance analysis was conducted to quantify the relative contribution of each input variable to the model’s predictive accuracy. Second, the SHapley Additive exPlanations (SHAP) method was employed to determine both the magnitude and direction of each feature’s influence on the model’s output.

The SHAP analysis provided a more granular understanding of how individual clinical and lifestyle features affected neuropathy severity predictions.

Before applying SMOTE, the dataset exhibited a pronounced imbalance, with a higher prevalence of non-neuropathic cases. Following preprocessing (sorting, imputation, encoding, and outlier removal), SMOTE was implemented to generate synthetic samples for minority classes, achieving a balanced class distribution as shown in Table 1.

Balancing the dataset significantly improved the model’s ability to identify minority-class patterns and enhanced predictive fairness across severity levels. Post-balancing analysis highlighted that sensory and neurological variables—particularly the tuning fork test, reflexes, and BMI—were the most discriminative indicators of neuropathy progression.

The initial dataset contained over 70 attributes. Features with excessive missingness (>80%) were removed, reducing the set to 20. Feature importance was then evaluated using RF and XGBoost algorithms, which provided embedded measures of variable relevance based on impurity reduction and information gain. Consistent results across both methods confirmed the reliability of selected predictors.

BMI and neurological indices (reflexes, tuning fork) emerged as dominant predictors, aligning with clinical evidence linking metabolic dysfunction and neural impairment to DPN progression.

4.4|eXtreme Gradient Boosting Feature Importance

XGBoost results were highly consistent with RF, confirming the stability of neurological and metabolic predictors in defining neuropathy severity.

These eight features collectively captured the clinical, neurological, behavioral, and genetic dimensions of DPN progression.

All composite (engineered) features were initially designed and tested. After evaluation through correlation (Pearson, Spearman), ANOVA, and feature importance analysis, four were selected for the final model.

These engineered indicators enhanced model accuracy from ~65% to over 86%, demonstrating their value in representing multi-domain pathophysiological interactions.

The dataset was split into 80% for training and 20% for testing to ensure a reliable generalization capability. For final evaluation, 5-fold cross-validation was used — training the model on four folds and testing on the remaining one, repeated five times, with average results reported for robustness.

The study evaluated several ML algorithms, including RF, XGBoost, LightGBM, SVM, MLP, and Voting Classifier, using Accuracy, Sensitivity, Specificity, F1-Score, and ROC-AUC as metrics.

I. RF achieved the highest accuracy (0.861) and F1-score (0.861), along with excellent sensitivity (0.859) and specificity (0.864).

II. XGBoost performed comparably (ROC-AUC = 0.845), maintaining strong balance across metrics. III. LightGBM achieved the highest ROC-AUC (0.846), showing great discrimination capability.

IV. SVM and MLP showed moderate but consistent results, with non-linear adaptability in MLP reflected in ROC-AUC (0.843).

V. Voting Classifier integrated multiple model strengths with ROC-AUC of 0.845, though slightly lower overall accuracy (0.821).

Based on the comparative results presented in Table 6, the RF model demonstrated the strongest overall predictive performance and was therefore selected as the final classification model. Feature importance analysis was subsequently conducted using the RF model to identify the most influential variables contributing to the classification task.

To further enhance model interpretability, SHAP analysis was additionally performed using the XGBoost model, as it provides a stable and well-established framework for SHAP-based explanations. The SHAP findings were consistent with the feature importance patterns obtained from the RF model and were used as a complementary interpretability approach, rather than to redefine the final predictive model.

RF and XGBoost models consistently identified BMI, tuning fork (vibration) test, and Mean Reflex as the most influential predictors of DPN severity. In the RF model, BMI exhibited the highest feature importance, followed by the tuning fork test and composite sensory indices, emphasizing the combined contribution of metabolic status and clinical–sensory measurements.

To investigate the contribution and direction of each feature, SHAP were applied; due to variability in RF SHAP values, the main interpretability analysis was conducted using XGBoost. SHAP results showed that lower Mean Reflex values were strongly associated with higher neuropathy severity, the tuning fork test effectively distinguished mild from severe cases, and higher BMI was directly associated with increased risk of severe DPN.

Additional composite indices (Neuro Index and Sensory Index) demonstrated clinically consistent effects, whereas demographic variables had minimal influence. The agreement between SHAP findings and RF feature importance confirms the robustness, interpretability, and clinical coherence of the proposed models.

In recent years, several studies have applied ML techniques to predict DPN or related diabetic complications. For instance, a recent systematic review reported that most ML-based DPN prediction models rely primarily on clinical and biochemical parameters [7]. Another study employed ML models to predict Diabetic Sensorimotor Polyneuropathy (DSPN) and concluded that algorithms such as RF perform effectively, particularly when electrophysiological or biochemical features are incorporated [14]. Moreover, a recent risk-prediction model based on laboratory and clinical data demonstrated the feasibility of ML-based approaches for DPN prediction in real-world clinical settings [3].

Compared with these earlier works, the present study introduces two key methodological improvements. First, a wider and more diverse feature set was utilized. While many previous studies focused mainly on biochemical markers (e.g., HbA₁c, serum glucose) or standard clinical and demographic risk factors, our model integrates neurological examination outcomes (reflex measurements and tuning-fork vibration test), anthropometric data (BMI), lifestyle and demographic variables (education and gender), and family history. Feature-importance analysis using RF and SHAP-based interpretability revealed that neurological and metabolic indicators—particularly Mean Reflex, vibration perception, and BMI—play a dominant role in predicting DPN severity, underscoring the clinical relevance of the selected features.

Second, a multi-level severity classification strategy was adopted. Instead of a binary classification (presence vs. absence of neuropathy), patients were categorized into multiple severity levels (none, mild, moderate, and severe). SHAP analysis further supported this approach by demonstrating how key features contribute differently across severity levels, providing more clinically actionable insights and better reflecting the progressive nature of DPN.

These methodological enhancements likely explain the improved predictive performance of the proposed model (Accuracy = 0.861, AUC = 0.855) and the consistent identification of neurological, metabolic, and demographic variables as jointly significant predictors.