Cost-Sensitive Gradient Boosting for High-Risk Clinical Diagnostics: Meticulous Hyperparameter Mapping of Custom Loss Functions for Asymmetric Medical Risk in Diabetes Detection

This research presents a cost-sensitive XGBoost framework for clinical diabetes detection that prioritizes the reduction of false negative predictions in high-risk medical screening. Traditional machine learning models optimize overall accuracy and treat all classification errors equally, which can be problematic in healthcare applications where missing a diabetic patient may have severe consequences. The proposed approach introduces a custom cost-sensitive loss function that applies higher penalties to false negative predictions. Experiments were conducted on the Pima Indians Diabetes Dataset using multiple penalty-weight configurations. Results demonstrate a significant reduction in missed diabetic cases and substantial improvement in diabetic recall compared to the standard XGBoost baseline. The study identifies an optimal penalty configuration that reduces clinical failures by 63.2% and improves diabetic recall from 65% to 87%. The findings highlight the importance of asymmetric error handling in medical AI systems and provide a practical framework for deploying machine learning models in healthcare screening, clinical decision support, and other high-risk diagnostic environments.