Addressing Data Imbalanced Issue for Thyroid Disease Prediction Using Hybrid Optimized Generative Adversarial Network

Abstract

Thyroid disease is a global health concern caused by the thyroid gland's inability to produce essential hormones. Early prediction of thyroid diseases is crucial to minimize severe effects. In the medical field, various Artificial Intelligence (AI) models have been developed to identify and classify thyroid diseases. The Conditional Tabular Generative Adversarial Network (CTGAN) model has been used to generate synthetic data and balance class distribution. However, it may face challenges such as mode collapse, where the discriminator is stuck in a local minimum, making CTGAN training difficult, especially with large datasets. Hence, a new Hybrid Optimized GAN (HOGAN) model is introduced in this paper to handle imbalanced data in thyroid disorder prediction and classification. The HOGAN model combines the Variational Auto-Encoder (VAE) and Optimized CTGAN (OCTGAN) models to combat mode collapse and generate data for the minority class by learning accurate class conditioning in the latent space. VAE is first used for feature learning from the majority classes to create a meaningful latent representation. This representation is then fed into OCTGAN to generate specific data for the minority disease class and rebalance the thyroid disease dataset. The OCTGAN develops a population of generators to engage in the adversarial game with the discriminator, resulting in robust GAN training and enhanced generator efficiency. It also incorporates Kullback-Leibler (KL) divergence loss and Wasserstein gradient penalty loss to improve convergence speed and prevent overfitting. The resulting balanced thyroid disease dataset is utilized to train various classifiers like Support Vector Machine (SVM), Random Forest (RF), Convolutional Neural Network (CNN) and Bidirectional Gated Recurrent Unit (BiGRU) for accurate disease classification. On the thyroid disease dataset from the University of California, Irvine Irvine (UCI), experimental findings show that the BiGRU model obtains remarkable results of 99.75% for all metrics, including accuracy, precision, recall, and F1-score. Similarly, using the Kaggle dataset for thyroid disease, BiGRU achieves 99.34% accuracy, 99.35% precision, 99.35% recall, and 99.34% F1-score.For both datasets, the suggested model outperforms the SVM, RF, and CNN models.