Explainable Machine Learning-Based Diabetes Risk Assessment Using Balanced CatBoost, Threshold Optimization and SHAP with Multilingual Accessibility

A significant example of using machine learning is diabetes risk assessment using routinely collected health indicators. The current screening methods may be ineffective due to class imbalance, complex interactions between risk factors, and limited availability. Explainability of predictive models. This study brings forward an explainable binary diabetes–risk assessment pipeline based on the 2015 Diabetes Health Indicators data are from the CDC Behavioral Risk Factor Surveillance System (BRFSS). The original this dataset consists of 22 columns (21 predictors plus the target column) and 253,680 records. Diabetes_binary. The workflow that was implemented considers female participants aged between 6-12 (Age codes 6-12). Approximately 45-79 years of the standard BRFSS age coding, eliminating duplicate records and assessing Logistic The models used are Regression, Random Forest, XGBoost, CatBoost, and a model with class weighted Balanced CatBoost. After duplicate If it is removed, the study cohort consists of 91981 observations,76643 are negative and 15338 are positive observations. Because the CatBoost is trained weights[1, 5] on the classes and the positive class is significantly smaller. Then threshold analysis of the range of 0.30 to 0.60 was performed to check the precision-recall trade-off. The resulting threshold value is 0.60, and Balanced is the classification for this classification. CatBoost obtains 77.66% accuracy, 39.75% precision, 65.87% recall, 49.58% F1-score, and 81.55% ROC-AUC. SHAP to interpret the final model, an analysis is performed on it, which involves the factors of General Health, BMI, High Blood Pressure, High Cholesterol, Age is one of the most important characteristics. A multilingual presentation layer is created offering English, Marathi and Hindi. Prediction messages without any change in the model. The study exemplifies an actual marriage of the practical and the possible. Imbalance aware boosting, threshold selection, explainability, and multilingual accessibility for research oriented diabetes risk assessment. Formal demographic fairness metrics, external validation, calibration analysis, and clinical no validation was undertaken and this is an important direction of future work..