Sensitivity-First Deep Learning Framework for Early Cardiovascular Disease Risk Screening

Abstract

Cardiovascular disease is the leading cause of death worldwide, however, many patients at high risk of cardiovascular disease are never identified until it is too late. The current risk calculators are based on simplistic and linear models with few features that are designed to be accurate for the greater majority and less sensitive, that is, they miss many high-risk patients. In this paper, we use the Cleveland Heart Disease and Framingham Heart Study datasets to train a hybrid wide-and-deep neural network that is specifically designed to detect the missed cases. We use missing value imputation, min-max scaling, and SMOTE balancing as the preprocessing steps. We also built in SHAP based explanations and an uncertainty aware routing module to the model so that it can aid the clinical decision-making process. Our model achieves an accuracy of 93.8, outperforming logistic regression, random forest, XGBoost and deep learning baseline models on all metrics. We have designed this framework to be as sensitive as possible, because in a preventive screening setting missing a high-risk patient is unacceptable.