Development and validation of a hypertension risk prediction model using electronic health records data from general practices in Hampshire and Isle of Wight integrated care system

BACKGROUND: Hypertension is a major risk factor for heart, stroke and kidney disease. Identifying patients at risk of developing hypertension early and applying preventative measures can reduce the disease burden and improve outcomes. This study aimed to develop and validate a hypertension risk prediction model utilising data from electronic health records in Hampshire and Isle of Wight Integrated Care System. METHODS: We conducted a retrospective study using data from the Oracle Health Data Intelligence platform. We used an observation period of all history up to 31st July 2017 and a prediction period in the following five years, up to 31st July 2022. We included all adult patients registered with a general practitioner without existing hypertension and other cardiovascular disease. We considered a total of 54 predictors and used feature selection based on a combination of random forest feature importance and feature significance in backwards stepwise logistic regression. The outcome was a new development of hypertension within the 5-year prediction period. We evaluated three types of predictive models, logistic regression, decision tree and random forest. The hypertension prediction model was further redeveloped by applying the same methodology to a different geographic population of patients registered to practices within Lewisham and Greenwich. RESULTS: We included 569,405 patients, and of those 58,833 (10.3%) developed hypertension within five years. A total of 39 predictive factors were included following feature selection. The logistic regression model slightly outperformed the other two models and achieved a ROC-AUC of 0.82 within both the training and testing cohort. The model achieved a sensitivity of 75.82% and specificity of 73.67% within the training cohort. Additional model on new geographic population achieved a ROC-AUC of 0.82 within the training cohort and 0.83 within the testing cohort. CONCLUSIONS: The model was developed using a substantially larger number of patients compared with existing models and demonstrated good performance. It is used to identify persons at high risk of developing hypertension and support the provision of prophylactic interventions with model outputs presented as a percentage risk score within population health management tools. Further research could include an assessment of the impact of this model in clinical practice.