AI-Driven risk prediction models for hypertensive emergencies in diabetic patients: validation in multi-ethnic cohorts

The study was undertaken to confirm artificial intelligence (AI) risk models of hypertensive emergencies among diabetic patients in multiethnic populations. The study was a multicenter historical cohort involving 24,718 diabetic and hypertensive patients from various ethnic groups (European, African, South Asian, Hispanic, East Asian, and Middle Eastern). The performances of three machine learning algorithms (XGBoost, neural network, and random forest) were contrasted with logistic regression. The outcomes showed that the XGBoost model, which recorded AUC values of 0.89 for Cohort B and 0.85 for Cohort B, was significantly better compared to standard models and had a high ability to identify evolving patterns such as systolic blood pressure fluctuation and kidney function changes. However, subgroup analyses revealed significant ethnic differences in model performance: sensitivity was lower in African-American (76.2%) compared to South Asian (88.1%) patients, and positive predictive value was 15% lower in Hispanics compared with East Asians. Additionally, poor calibration in high-risk groups (African-Americans) and the influence of social determinants of health on predictive accuracy were observed. These findings reaffirm the importance of validating models in every ethnic environment, including social variables, and developing dynamic calibration procedures to provide equitable and accurate treatment.