Risk stratification of impaired glycemic control remains a major challenge in biomedical data analysis due to heterogeneous metabolic, behavioral, and therapeutic factors observed in large-scale populations. This study proposes a calibrated and interpretable decision–support framework, termed Calibrated Multi-Task Stacking Ensemble (CMSE), for joint modeling of clinically related glycemic outcomes. The framework integrates demographic variables, lipid profiles, renal and inflammatory biomarkers, dietary and smoking indicators, and therapy-related features within a unified predictive architecture. Robust modeling is ensured through leakage-aware preprocessing, quantile-based Winsorization, out-of-fold stacking, and isotonic calibration of probabilistic outputs. The physiological coherence between short-term and long-term glycemic markers is investigated using an explicit intertask coupling mechanism based on the estimated average glucose (eAG) ratio. Model interpretability is supported using SHAP analysis, mutual information, distance correlation, and feature importance metrics. In the primary medication-free screening configuration, the framework is evaluated on the NHANES 2017–March 2020 dataset, achieving ROC-AUC of 0.865 for diabetes classification and R² values of 0.385 and 0.366 for plasma glucose and HbA1c prediction, respectively. These results indicate that CMSE provides a reliable and explainable approach for calibrated glycemic risk assessment and clinical decision support.

This study presents NORIA, an availability-aware multitask learning framework for clinical prediction from incomplete electronic health records (EHRs). We analyzed a public dataset of 77,724 patients and 107 numerical features. To prevent temporal leakage, predictors were grouped by availability, and training-only feature selection identified three baseline and two first-year variables. NORIA jointly predicts baseline and updated glycemic control, an HbA1c change index, and 21 medication labels using shared representations, masked training, and stage-specific prediction updates. Comparators included classical machine learning, gradient boosting, deep tabular models, and multitask architectures (MMoE and PLE). On the internal test set, NORIA achieved AUROCs of 0.8554 and 0.8888 for baseline and updated glycemic control, respectively, and medication macro-AP of 0.1854 versus 0.1879 for CatBoost. The predefined medication improvement criterion was not met. With 20% of baseline features artificially masked, updated-control AUROC declined by 0.0271 for NORIA, versus 0.0510 for CatBoost and 0.0542 for FT-Transformer. Ablation analyses indicated that masked training provided the clearest robustness benefit. NORIA thus demonstrated competitive accuracy and robustness in selected simulated missing-data scenarios, without establishing overall superiority.