Investigation on Machine Learning Models for Predicting Diabetes Risk in Indian Populations

Abstract

Background: Diabetes mellitus is a major public health concern in India, with increasing prevalence driven by demographic, metabolic, behavioral, and lifestyle-related factors. Early identification of individuals at high risk of diabetes can support timely prevention and improve health outcomes. Machine learning (ML) approaches offer opportunities to identify complex and nonlinear relationships among multiple risk factors and may complement conventional statistical approaches. Objective: This study aimed to investigate and compare the performance of different ML models for predicting diabetes risk, with particular emphasis on identifying the most influential predictors and assessing the potential applicability of ML-based approaches for early risk stratification in Indian populations. Methods: A cross-sectional analytical approach was used to evaluate demographic, physiological, and lifestyle-related variables associated with diabetes risk. The study considered variables including age, body mass index (BMI), blood glucose, blood pressure, insulin, family history of diabetes, and physical activity. Data preprocessing included missing-value management, outlier identification, categorical encoding, and feature standardization. The dataset was divided into training (70%) and testing (30%) subsets using stratified sampling. Seven supervised ML algorithms—Logistic Regression, Decision Tree, Random Forest, Support Vector Machine, K-Nearest Neighbors, Extreme Gradient Boosting (XGBoost), and Artificial Neural Network—were compared. Model performance was assessed using accuracy, precision, recall, F1-score, and area under the receiver operating characteristic curve (AUC-ROC). Pearson correlation, one-way ANOVA, and multivariate logistic regression were additionally used to examine associations between predictors and diabetes risk. Results: XGBoost demonstrated the strongest overall predictive performance, achieving an accuracy of 89.2%, precision of 0.88, recall of 0.87, F1-score of 0.87, and AUC of 0.93. Random Forest and the neural network also demonstrated strong performance, with accuracy of 87.6% and 88.5% and AUCs of 0.91 and 0.92, respectively. Glucose level was the most influential predictor, followed by BMI and age. Statistical analyses further supported the importance of these metabolic and demographic factors in diabetes risk prediction. Conclusion: Ensemble and advanced ML approaches, particularly XGBoost and Random Forest, demonstrated promising performance for diabetes risk prediction. Integrating ML with statistical analysis and interpretable AI approaches may strengthen early risk identification and support evidence-based diabetes prevention and personalized healthcare strategies in Indian populations. Further validation using larger, diverse, and multi center datasets is required before clinical implementation.