EvoX-Boost: a hybrid evolutionary GPU-XGBoost framework for multiclass thyroid disease classification

Abstract

Timely and correct diagnosis of the thyroid disease is crucial for thyroid metabolic and cardiovascular complications. However, real-world EMR data are multivariable, heterogeneous and multi-class, which makes traditional ML models sensitive to hyperparameter configurations and computationally prohibitive when considering public multiclass thyroid data. In this paper, to cope with the challenges, we propose EvoX-Boost, a hybrid evolutionary optimization scheme for GPU-based XGBoost for EMR-like, high dimensional thyroid classification. EvoX-Boost integrates (i) Bayesian optimization with Optuna serving as a rough global search, (ii) real-space differential evolution (DE) being used as a refinement step, (iii) a GA-DE hybrid phase for enhanced exploitation and convergence stability. The goal of this architecture is to scale with CUDA parallel training. Experiments we carried out on real 15-class thyroid dataset(22, 632 records) demonstrate that EvoX-Boost gets 99.16% as the best fold accuracy and 98.90% as the mean accuracy over 5-fold cross validation. Besides, the macro-averaged precision, recall, F1-score are 73.67%, 63.12%, 66.16% respectively. Comprehensive evaluation, including confusion matrices, ROC curves, feature importance analysis, and runtime profiling, demonstrates that the proposed framework achieves strong predictive performance and shows potential for future clinical decision-support applications.