The chronic metabolic disorder known as type II diabetes mellitus (T2DM) is an enormous global health concern. In order to create an open and reliable predictive model, this study integrated framework that combines explainable artificial intelligence (XAI), chi-square statistical validation, and genetic algorithms (GA) as primary tool for selecting significant features. To extract trustworthy genomic features, whole-exome sequencing (WES) data from the Mizo population was preprocessed and examined. An ideal subset of 23 features was found using GA, and their statistical significance was confirmed through additional validation using chi-square tests. With an accuracy of 86.5%, F1-score of 0.8708, and recall of 0.91 corresponding to Optuna-based hyperparameter optimisation, CatBoost performed better than LightGBM among the different classifiers. SHapley Additive exPlanations (SHAP) beeswarm and force plots were used to make the model interpretable, exposing both instance-level and feature contributions. The findings show that combining statistical validation with evolutionary feature selection enhances clinical interpretability and predictive accuracy. This framework offers important insights for precision medicine and data-driven healthcare by helping to create explainable, statistically sound, and generalisable predictive models for T2DM.