Improving Cardiovascular Disease Forecasting with Machine Learning and Electronic Medical Record Data Characteristics Within a Local Healthcare Network

Abstract

The PCE Risk Calculator, developed by the ACC/AHA, is frequently utilized in the United States for the purpose of averting the onset of Atherosclerotic cardiovascular disease (ASCVD) via first-line defense strategies. However, this calculator may not accurately estimate risk for certain populations, potentially leading to either under- or over-estimation of risk. We have created calculator for ASCVD risk specific to a population by leveraging advanced Machine Learning (ML) techniques and Electronic Medical Record (EMR) data. Our study involved comparing predictive accuracy of our calculator with PCE calculator. Between January 1, 2009, and April 30, 2020, data was gathered from 101,110 distinct EMRs of patients who were actively receiving treatment. Patient datasets underwent machine learning techniques containing Longitudinal (LT) and Cross-Sectional (CS) features, or solely CS features, derived from laboratory values and vital statistics. The models' effectiveness was assessed using fresh price metric (Screened Cases Percentage @Sensitivity level). In terms of prediction accuracy, every ML model that was tested performed better than the PCE risk calculator. Area Under Curve (AUC) score of 0.902 was obtained by Random Forest (RF) ML technique when CS and LT characteristics were combined (RF-LTC). Our machine learning model only needed to screen 43% of patients in order to identify 90% of positive ASCVD cases, in contrast to the PCE risk calculator, which required screening 69% of patients. Prediction models created using ML techniques reduce the amount number of tests necessary to forecast ASCVD and increase the accuracy of ASCVD prediction when compared to using PCE calculator alone. The combination of LT and CS features in these ML models leads to a significant enhancement in comparing the ASCVD prediction to utilizing CS features exclusively.