Development of a Machine Learning Model and Explainable Artificial Intelligence for Hypertension Risk Estimation System

Abstract

Hypertension is a leading modifiable risk factor for cardiovascular disease, yet its management remains largely reactive, so many cases are detected only after complications arise. Machine learning (ML) can estimate hypertension risk earlier, but the black-box nature of high-performing models hinders clinical trust; furthermore, the numerical outputs of Explainable Artificial Intelligence (XAI) are difficult for laypeople to interpret, and generic Large Language Models (LLMs) are prone to medical hallucination. This study designs and implements a hypertension risk screening system that integrates ML, XAI, and an LLM so that predictions are not only accurate but also transparent and communicative. Following the CRISP-DM framework on secondary data from the fifth wave of the Indonesian Family Life Survey (IFLS-5), data acquisition focused on three modules (B3A, B3B, and BUS) and yielded 30,251 adult respondents described by eleven non-invasive features. Four algorithms—Logistic Regression, Random Forest, XGBoost, and CatBoost—were compared through a two-scenario ablation study, and CatBoost was selected as the final model, achieving a ROC-AUC of 0.78 and an accuracy of 0.71 at the Youden Index operating point, thereby satisfying the predefined success criteria of ROC-AUC ≥ 0.75 and accuracy ≥ 0.70. Model interpretation employed SHAP (SHapley Additive exPlanations), while the explanatory narrative was built through a two-layer architecture: all numbers, contribution directions, and facts are fixed deterministically from the SHAP values, and the LLM (llama-3.1-8b-instant) only simplifies the language for lay readers, without any external retrieval (RAG), thereby suppressing the risk of hallucination. All components were integrated into a web-based prototype. Validation by ten physicians over five representative cases yielded an overall mean of 3.45 on a 1-5 Likert scale (valid), with risk-status appropriateness and narrative factuality rated valid and the SHAP contribution-direction and magnitude rated adequately valid. A field test with 53 lay users further indicated high clinical usefulness and acceptance, with mean scores above 4.1. The system therefore demonstrates the ability to deliver hypertension risk estimates that are accurate, transparent, and easy to understand, suitable as an early-screening and risk-factor-education aid rather than a substitute for professional medical diagnosis.