HybGANN as an interpretable generative evolutionary model for predicting diabetes from imbalanced data

The digitisation of medical data has enabled large-scale analysis. However, clinical datasets, such as those used for diabetes prediction, often have class imbalances, with disease cases significantly underrepresented. This imbalance poses a significant challenge for traditional machine learning models, which tend to favour the majority classes. Additionally, many high-performance models operate as black boxes, which limits their adoption in clinical practice due to their lack of interpretability. This paper introduces HybGANN, a hybrid framework that integrates Conditional Tabular Generative Adversarial Network (CTGAN) for synthetic minority data generation, a unique hybrid genetic algorithm (GA) that co-evolves hyperparameters and internal weights from artificial neural networks (ANNs) in a Lamarckian fashion, and SHapley Additive Explanations (SHAP) for post-hoc model interpretability. HybGANN provides a semi-automated workflow that improves predictive performance while ensuring transparency and adaptability. Applied to a large-scale diabetes dataset, experiments have demonstrated that the HybGANN model outperforms a benchmark ANN network that also uses the same CTGAN pre-balanced dataset on all key classification metrics. The framework achieves ROC-AUC and PR-AUC values of 0.9184 and 0.9268, respectively, demonstrating its effectiveness and potential as a reliable AI solution for clinical decision support in imbalanced medical fields.