Enhanced Diabetes Detection Using Generative Adversarial Network Techniques

For effective treatment and management of Type 2 Diabetes Mellitus (T2DM), precise and timely diagnosis is of utmost importance. Nonetheless, clinical data that is too sparse or unstructured often manipulates the automated diagnostic systems and adversely impact the prediction accuracy of T2DM clinical algorithms. In this work, the problem is addressed by developing sythetic patient data through Generative Adversarial Network (GAN) approaches and proposes a new method for diabetes identification which improves dataset augmentation and diversity. Ensemble classifers, with emphasis on Random Forests, are orthogonal primary focus trained on the optimized set. Models trained on data augmented from GAN outperformed models that were trained on only original datasets according to experimental results, achieving improvement across AUC score, sensitivity, specificity, and overall accuracy. Predictive AUC scores were enhanced when sensitive classifiers biased towards T2DM positive samples were trained. The disclosed results illustrate the potential of using data augmentation as a GAN-based technique to improve model performance and assist in the diabetes screening clinical workflow.