Improving Cardiovascular Disease Risk Prediction With Machine Learning Using Mental Health Data: A Prospective Uk Biobank Study

Dorraki,M.,Liao,Z.,Abbott,D.,Psaltis,P.,Baker,E.,Bidargaddi,N.,Wardill,H. R.,Van Den Hengel,A.,Narula,J.,Verjans,J. W.

DOI: https://doi.org/10.1101/2022.10.23.22281428

2022-10-26

MedRxiv

Abstract:Background: Robust and accurate prediction of cardiovascular disease (CVD) risk facilitates early intervention to benefit patients. It is well-known that mental disorders and CVD are interrelated. Nevertheless, psychological factors are not considered in existing models, which use either a limited number of clinical and lifestyle factors, or have been developed on restricted population subsets. Objectives: To assess whether inclusion of psychological data could improve CVD risk prediction in a new machine learning (ML) approach. Methods: Using a comprehensive, long-term UK Biobank dataset (n=375,145), we examined the correlation between CVD diagnoses and traditional and psychological risk factors. An ensemble ML model containing five constituent algorithms [decision tree, random forest, XGBoost, support vector machine (SVM), and deep neural network (DNN)] was tested for its ability to predict CVD risk based on two training datasets: one using traditional CVD risk factors alone, or a combination of traditional and psychological risk factors. Results: Our ensemble ML model could predict CVD with 71.31% accuracy using traditional CVD risk factors alone. However, by adding psychological factors to the training data, accuracy dramatically increased to 85.13%. The accuracy and robustness of our ensemble ML model outperformed all five constituent learning algorithms. Re-testing the model on a control dataset to predict bone diseases returned random results, confirming specificity of the training data for prediction of CVD. Conclusions: Incorporating mental health assessment data within an ensemble ML model results in a significantly improved, highly accurate, state-of-the-art CVD risk prediction.