Lifetime vs 10-year Cardiovascular Disease Prediction in Young Adults Using Statistical Machine Learning and Deep Learning: The CARDIA Study

Abstract

ABSTRACTImportanceNational guidelines for primary prevention suggest consideration of lifetime risk for cardiovascular (CV) disease in addition to 10-year risk, however, it is unclear if the predictors of 10-year vs lifetime (10-26 years) CV events are similar.ObjectiveTo use a combination of machine learning methods with deep phenotyping to differentiate 10-year versus lifetime predictors of CV outcomes.Design, Setting, and ParticipantsThis retrospective analysis used the prospectively collected data from the CARDIA (Coronary Artery Risk Development in Young Adults) study, a cohort of White and Black participants recruited from four clinical centers in the US. The analysis included 4314 participants, aged 23-35 years who were then followed up over 25 years through August 2018.Main Outcomes and Measures449 variables collected in 1990-91 from imaging and noninvasive tests, questionnaires, and biomarker panels were included. We used machine learning techniques to identify the top-20 predictors of both 10-year and lifetime (10-26 years) CV events (coronary heart disease, myocardial infarction, acute coronary syndrome, stroke, transient ischemic attack, heart failure, peripheral arterial disease, and CV death).ResultsKidney disease, family history of CV disease, and echocardiographic parameters of left ventricular systolic and diastolic dysfunction, and hypertrophy were important markers of 10-year CV events. Traditional risk factors and indices of body size featured heavily as top predictors of lifetime CV risk. Among the different machine learning techniques, Random Survival Forest and Nnet-survival performed the best (C-index of 0.80 for 10-year and 0.72 for lifetime). These models outperformed Cox models including traditional CV risk factors.Conclusions and RelevanceFamily history of CVD, kidney disease, and subclinical phenotyping of CVD using echocardiography are important for 10-year risk estimation. However, traditional CV risk factors alone may be adequate in estimating lifetime CV risk.Key PointsQuestionDo machine learning (ML) and deep learning (DL)-based survival analysis models help differentiate 10-year versus lifetime predictors of cardiovascular (CV) outcomes in young adults?FindingsIn this retrospective analysis of 4314 participants in the CARDIA study, ML and DL survival analysis improved CVD risk prediction over traditional Cox models and revealed the top 20 predictors among 449 variables. Top 10-year risk predictors include kidney disease, family history of CV disease, and echocardiographic parameters, where as traditional risk factors and indices of body size featured heavily as top predictors of lifetime CV risk.MeaningFamily history, kidney disease, and subclinical phenotyping of CVD using echocardiography play a prominent role for 10-year risk estimation, while traditional CV risk factors alone may be adequate in estimating lifetime CV risk in young adults.