Early Detection of Type 2 Diabetes Using Machine Learning: A Fairness-Aware Study for Underserved Populations

Type 2 diabetes mellitus represents one of the most urgent public health challenges in sub-Saharan Africa, where late diagnosis continues to drive preventable complications. This study presents a comprehensive, fairness-aware comparative evaluation of seven machine learning algorithms for the early prediction of Type 2 diabetes, using the Pima Indians Diabetes Database and validated against the CDC Behavioral Risk Factor Surveillance System dataset. The evaluated models include Logistic Regression, Decision Tree, Random Forest, Support Vector Machine, K-Nearest Neighbors, XGBoost, and a Multi-Layer Perceptron. A rigorous preprocessing pipeline incorporating median imputation, SMOTE oversampling, and feature normalization was applied. Random Forest achieved the highest AUC-ROC of 0.8226, while XGBoost recorded the strongest cross-validated accuracy at 81.12 percent. SHAP analysis identified glucose concentration and BMI as the dominant predictive features. A fairness analysis across age and BMI subgroups revealed significant performance disparities, with critical implications for equitable AI deployment in low-resource clinical settings. The findings directly inform the development of Nexora's AI-driven health intelligence platform for underserved communities across Ghana and West Africa. This work forms the scientific foundation of an ongoing research program to build clinically actionable, fair, and interpretable machine learning tools for preventive healthcare in emerging economies.