DEEP LEARNING MODELS FOR PREDICTING THE RISK OF CARDIOVASCULAR INCIDENTS BASED ON THE WISCONSIN LONGITUDINAL STUDY

Abstract

Objective: Existing models use only a limited number of well-known factors such as age, sex, cholesterol level, glucose level, and heart rate, while there is no general pipeline, which could give better overall prediction power, for an automatic pooling of the most correlated with heart disease predictors from the large-scale datasets. They often require preprocessing, and dimensionality reduction and has many missing data. Design and method: To address such challenges, we implemented data processing pipeline and several predictive models based on the Wisconsin Longitudinal Study dataset, which has both epidemiological and genetic data. We used univariate Cox Proportional Hazard models to estimate the effect of each variable, Hyperimpute to impute missing data, DeepSurv neural network and Autoprognosis 2.0 to develop final models. Fig. 1 The research pipeline, where MI – Myocardial Infarction, HA – Heart Attack. Results: The pipeline showed high predictive power (approx. 80% under the ROC curves) to predict risks of heart incidents such as heart attack, myocardial infarction, and stroke within a 5–10-year time window. Furthermore, we estimated the effect of various polygenic risk scores like subjective well-being, neuroticism, and depression on heart incidents. Conclusions: We implemented an effective pipeline of accurate model development to predict risks of heart incidents, which, as a result, can be used to build different models to estimate other risks correlated with heart incidents and could be used on other biobanks’ datasets. Therefore, this opens more ways to explore the causalities of heart diseases better and improve treatments of such conditions.