Generative adversarial networks and hyperparameter-optimized XGBoost for enhanced heart disease prediction

Abstract

Heart Disease (HD) serves as the primary reason for worldwide deaths which creates an urgent requirement to create diagnostic intelligent systems that can identify patients at their earliest stage. The development of strong predictive models faces a significant obstacle because clinical data contains both biased information and noisy elements which create unreliable results. This study proposed a hybrid framework, GAN-XO, for early HD detection. The methodology addressed fundamental data quality issues through a multi-step preprocessing pipeline. First, a Generative Adversarial Network (GAN) is employed for synthetic oversampling of the minority class, balancing the imbalanced PKIOHD and Framingham datasets. Second, a two-phase outlier detection and elimination process using z-score and Interquartile Range (IQR) methods is applied to obtain clean data. Finally, an XGBoost classifier, whose hyperparameters are optimally tuned using Optuna, is utilized for prediction. The designed GAN-XO system achieved its best accuracy score of 96.60% and F1-Score value of 0.9649 through testing on an evenly distributed and completely clean test dataset. The ablation research demonstrated that all system elements provided distinct benefits while the outlier elimination process turned out to be vital for improving model performance and trustworthiness. The research results demonstrate that accurate healthcare predictions require both integrated modelling systems and precise management of data quality. The substantial performance enhancement shows that the framework improves clinical decision-making while enabling correct diagnosis of HD.