•  
  •  
 

Mesopotamian Journal of Artificial Intelligence in Healthcare

Abstract

This study investigates the use of Genetic Algorithm-based hyperparameter optimization (GA-HP) for diabetes prediction with deep neural networks (DNNs). The proposed pipeline combines GA-HP with mutual information (MI) feature selection and SMOTE class balancing applied within training folds. The framework optimizes eight DNN hyperparameters using the validation area under the ROC curve (AUC) as the fitness function. Evaluation was performed using a held-out test set and stratified five-fold cross-validation on the Pima Indians Diabetes Dataset (PIDD) and the Early Stage Diabetes Risk Prediction Dataset (ESDRPD). Across three independent runs, the held-out test AUC and Matthews correlation coefficient (MCC), reported as mean ± standard deviation, were 0.938 ± 0.013 and 0.746 ± 0.007, respectively, for PIDD, and 0.993 ± 0.006 and 0.911 ± 0.025, respectively, for ESDRPD. The ablation analysis showed that adding GA-HP to the pipeline containing SMOTE and MI feature selection improved the mean AUC on both datasets, while increasing the number of false negatives on PIDD and reducing them on ESDRPD. These results indicate dataset-dependent gains and highlight the need to evaluate discrimination ability alongside sensitivity.

Share

COinS