Artificial Intelligence–Derived Risk Prediction: A Novel Risk Calculator Using Office and Ambulatory Blood Pressure

BACKGROUND: Quantification of total cardiovascular risk is essential for individualizing hypertension treatment. This study aimed to develop and validate a novel, machine-learning–derived model to predict cardiovascular mortality risk using office blood pressure (OBP) and ambulatory blood pressure (ABP). METHODS: The performance of the novel risk score was compared with existing risk scores, and the possibility of predicting ABP phenotypes utilizing clinical variables was assessed. Using data from 59 124 patients enrolled in the Spanish ABP Monitoring registry, machine-learning approaches (logistic regression, gradient-boosted decision trees, and deep neural networks) and stepwise forward feature selection were used. RESULTS: For the prediction of cardiovascular mortality, deep neural networks yielded the highest clinical performance. The novel mortality prediction models using OBP and ABP outperformed other risk scores. The area under the curve achieved by the novel approach, already when using OBP variables, was significantly higher when compared with the area under the curve of the Framingham risk score, Systemic Coronary Risk Estimation 2, and Atherosclerotic Cardiovascular Disease score. However, the prediction of cardiovascular mortality with ABP instead of OBP data significantly increased the area under the curve (0.870 versus 0.865; P =3.61×10 − 28 ), accuracy, and specificity, respectively. The prediction of ABP phenotypes (ie, white-coat, ambulatory, and masked hypertension) using clinical characteristics was limited. CONCLUSIONS: The receiver operating characteristic curves for cardiovascular mortality using ABP and OBP with deep neural network models outperformed all other risk metrics, indicating the potential for improving current risk scores by applying state-of-the-art machine learning approaches. The prediction of cardiovascular mortality using ABP data led to a significant increase in area under the curve and performance metrics.