Framingham risk score prediction at 12 months in the STANDFIRM randomised control trial

BackgroundThe Shared Team Approach between Nurses and Doctors For Improved Risk factor Management (STANDFIRM, ACTRN12608000166370) trial was designed to test the effectiveness of chronic disease care management for modifying the Framingham risk score (FRS) among patients with stroke or transient ischemic attack. The primary outcome of change in FRS between baseline and 12 months was not met. We aimed to determine characteristics of participants at baseline that predict reduction in FRS at 12 months and whether future FRS is predetermined at the time of randomization MethodData included 35 variables encompassing demographics, risk factors, psychological, social and education status, and laboratory tests. Five supervised machine learning (ML) methods were used: random forest (RF), extreme gradient boosting (XGBoost), support vector regression (SVR), multilayer perceptron artificial neural network (MLP) and K-nearest neighbor (KNN). We split data for training (80%, n=406) and testing (20%, n=102). ResultsTraining and test data were evenly matched for age, sex, baseline and 12-month FRS. Following tuning of the five ML methods, the optimal model for predicting FRS at 12 months was SVR (R2=0.763, root mean squared error or RMSE=8.52). The five most important variables for SVR were: baseline FRS, age, male sex, sodium/potassium excretion and proteinuria. All ML methods were poor at determining change in FRS at 12 months (R2<0.161). ConclusionOur findings suggest that change in FRS as an endpoint in trials may have limited value as it is largely determined at baseline. In this cohort, Support Vector Regression was the optimal method to predict future but not change in FRS.