An Interpretable Framework Based on Knowledge Distillation for the Hypertension Early Warning Model

For any machine learning or deep learning model, explainability is of paramount importance. As data complexity increases, deep learning models renowned for their outstanding performance have been applied to a wide range of tasks. However, these models remain typical “black boxes” whose predictions are difficult to interpret. To address this limitation, this study proposes a progressive framework that integrates feature engineering with knowledge distillation for interpreting deep learning-based hypertension prediction models. Statistical relevance testing are performed to select target-associated variables, and numeric variables are first partitioned with a supervised decision tree to obtain interval-based, clinically legible features with interpretable thresholds. On this basis, FT-Transformer is adopted as the teacher to generate soft labels that capture nonlinear structure for distillation. Finally, a globally interpretable linear model is trained on the transformed features and soft labels, preserving accuracy while delivering quantitative, interval-level risk attribution. Experiments on the public MIMIC-III demonstrate that our method significantly outperforms traditional interpretable models and machine learning baselines. Although AUC and accuracy slightly lag behind deep learning models, our approach delivers intuitive feature interpretations while maintaining competitive performance.