A reliable and explainable deep learning framework for clinical-grade endocrine disorder risk prediction and decision...

The reliable prediction of endocrine disorders remains a significant challenge in clinical decision support, particularly regarding the simultaneous achievement of predictive accuracy, interpretability, and reliability. In this study, we propose a Gated Multi-Task Attention Network (GMTAN) for endocrine disorder prediction and clinically oriented decision support. The framework jointly models thyroid disorder and polycystic ovary syndrome (PCOS) prediction using heterogeneous endocrine datasets while integrating uncertainty estimation and explainability within a unified architecture. The proposed model combines shared feature representation learning with task-specific gating and attention mechanisms to capture both common and disease-specific endocrine patterns. To improve reliability, Monte Carlo Dropout is incorporated during inference to estimate predictive uncertainty and confidence. In addition, Shapley Additive exPlanations (SHAP)-based feature attribution and attention visualization are used to provide clinically interpretable explanations for individual predictions. Experiments were conducted using the UCI Thyroid dataset and the Kaggle PCOS clinical dataset. To improve reproducibility, all experiments were repeated across multiple runs using fixed random seeds, and model performance was evaluated using classification, calibration, and reliability metrics. The proposed GMTAN achieved area under the receiver operating characteristic curve (AUROC) scores of 0.98 and 0.97 for thyroid and PCOS prediction tasks, respectively, demonstrating improved calibration performance compared with baseline machine learning and deep learning models. The results suggest that integrating multi-task learning, uncertainty-aware inference, and explainability within a single framework can improve both predictive performance and interpretability for endocrine disorder prediction. While additional clinical validation is still necessary, the proposed framework demonstrates potential as a clinically oriented assistive decision support system.