Automated Prediction of Heart Disease based on EMR data using Deep Learning Algorithm: A Review and Evaluation Study for various Feature Extraction Techniques

Abstract

For many years, healthcare providers in the United States have relied on the ACC/AHA Pooled Cohort Equations (PCE) Risk Calculator as their main resource for estimating a person’s likelihood of developing atherosclerotic cardiovascular disease, a leading form of heart-related illness. Despite being widely used, the calculator occasionally overestimates or underestimates risk but doesn't always perform equally across different groups. To overcome these challenges, we developed an automated ASCVD risk-prediction model tailored to particular patient groups, using machine-learning techniques applied to real-world electronic medical record data. We then compared its performance with the PCE approach. In our study, we reviewed 101,110 electronic medical records from patients seen between January 1, 2009, and April 30, 2020. The machine-learning models were trained using either cross-sectional clinical data alone or a combination of cross-sectional information and longitudinal patterns drawn from lab results and vital-sign measurements. To understand how each model identified true cases, we introduced a cost-focused metric called the “Screened Cases Percentage at a given Sensitivity,” which shows how many patients would require follow-up testing to detect most ASCVD cases. In every analysis we conducted, the machine-learning models outperformed the PCE calculator. The strongest results came from a random forest model that used both cross-sectional and longitudinal data, achieving an AUC of 0.902 (95% CI: 0.895–0.910). To identify 90% of ASCVD cases, this approach required screening only 43% of patients, compared with 69% when relying on the PCE. Overall, combining CS and LT data led to accurate predictions and fewer unnecessary screenings.