Machine learning-based classification of diabetes mellitus using sociodemographic, behavioral, and clinical predictor

Diabetes mellitus, particularly type 2 diabetes, is a growing global health challenge with rising prevalence in low- and middle-income countries where access to diagnostic services is limited. Early detection is critical but traditional screening methods are often resource-intensive and inaccessible. Machine learning (ML) offers improved predictive accuracy by capturing complex interactions in health data, yet its use remains limited in resource-constrained settings. This study aims to develop ML models using population-level data to enhance diabetes risk prediction and support early intervention efforts. This study analyzed pooled data from WHO STEPS surveys (2017–2023) across 14 low- and lower-middle-income countries, comprising 56,560 adults (16,357 diabetic; 45,423 non-diabetic). Behavioral, clinical, anthropometric, and sociodemographic variables were preprocessed through one-hot or label encoding, with body mass index derived from height and weight. To address class imbalance, SMOTEENN was applied to the training data. Different machine learning algorithms including Logistic Regression, Decision Tree, KNN, SVM, Guassian NP, MLP, Gradient Boosting, AdaBoost, Random Forest, and XGBoost were trained on 80% of the data and evaluated on the held-out 20% using accuracy, balanced accuracy, precision, recall, F1-score, and AUC. Recursive Feature Elimination selected the top 10 predictors, and XGBoost performance was further examined using confusion matrices, ROC and precision–recall curves, while SHAP values were used to interpret feature contributions. Among the evaluated models, ensemble methods showed the strongest performance. XGBoost achieved the best overall results (accuracy = 0.90, balanced accuracy = 0.87, F1 = 0.93, AUC = 0.94), outperforming Random Forest, Gradient Boosting, and all traditional classifiers. MLP also performed well (accuracy = 0.89, AUC = 0.92), while logistic regression had the lowest recall. The XGBoost confusion matrix showed high discrimination, correctly identifying 2,586 diabetic and 8,549 non-diabetic individuals with relatively few misclassifications. Precision–recall analysis demonstrated consistently high precision (> 0.95 across most recall levels), indicating strong reliability in detecting true diabetes cases. SHAP interpretation revealed cholesterol, age, residency, tobacco use, and waist circumference as the most influential predictors of diabetes risk. Overall, the findings highlight the potential of machine learning to strengthen diabetes screening, early detection, and population-level surveillance, especially in resource-limited settings where access to diagnostic tools remains constrained. By integrating explainable ML methods with diverse contextual data, this study adds to the growing evidence supporting data-driven approaches for improving non-communicable disease prevention and control in low- and middle-income countries.