A Multi-Task Deep Learning Framework with Clinical Knowledge–Guided Regularization for Concurrent Prediction of Diabetes and Hypertension

The concurrent prevalence of diabetes and hypertension poses a major challenge in predictive healthcare, as both conditions share intertwined physiological and metabolic pathways. However, most existing predictive models treat them in isolation, overlooking their interdependencies. This study introduces a novel hybrid framework that combines domain knowledge–driven feature engineering, multi-task deep learning, and clinical knowledge–guided regularization to predict diabetes and hypertension concurrently. Utilizing the PIMA Diabetes and PPG-BP datasets, data preprocessing incorporated domain-aware imputation, outlier Winsorization, normalization, and Synthetic Minority Oversampling (SMOTE) to ensure balanced and reliable training data. A Multi-Task Feed-Forward Deep Neural Network (FFDNN) was designed to learn shared representations across related diseases while preserving task-specific distinctions. Additionally, a clinical regularizer was integrated into the loss function to constrain predictions within physiologically valid boundaries. Experimental results demonstrated superior performance, achieving 99.8% accuracy for diabetes, 95.5% for hypertension, and 96.0% for concurrent prediction. The framework significantly outperformed existing benchmarks, confirming that embedding clinical reasoning within deep learning models enhances robustness, interpretability, and diagnostic reliability, which are key requirements for real-world clinical deployment.