Optimizing Predictive Models for Healthcare: A Comparison of Personal and General Datasets in Coronary Artery Disease Risk Assessment

This research investigates how the choice of training population, personalized versus more general datasets, impacts the performance and generalizability of machine learning models for predicting Coronary Artery Disease risk. By comparing models trained on datasets of varying specificity using data from the UC Irvine Machine Learning Repository, the study explores trade offs between accuracy and broad applicability to improve strategies for early CAD risk assessment and prevention.

Despite the field of medicine continuing to grow, both in reach and technological advancement, Coronary Artery Disease (CAD) continues to be the leading cause of mortality in the United States [3]. CAD is a common heart condition that occurs when plaque buildup narrows the arteries, potentially leading to heart attacks and other complications that prevent the flow of blood. Assessing CAD risk is crucial yet challenging due to variability in patient data of different cardiovascular factors. Millions of people every year die of CAD, making early risk assessment a critical necessity. Machine learning has become a powerful tool in healthcare, demonstrating its potential to identify individuals at risk based on their emerging symptoms and other health factors, while providing targeted procedures to save lives. But we have realized that the effectiveness of these models depends highly on the datasets used to train them. Choosing the right training population for these models is a nuanced problem. Training a model on too specific a population risks not generalizing to a larger population, but training too large a population may risk not being personalized enough to individual patients. This paper aims to compare the efficacy of predictive models trained on more personalized datasets versus general datasets that source their data from wider and smaller populations, using data sourced from the UC Irvine Machine Learning Repository. After training models on different populations, we analyzed the results to see if there are trade-offs between accuracy and generalizability when using personal and general datasets to predict CAD. By contributing to the further optimization of predictive models for CAD, we hope to explore more effective strategies to increase CAD risk prevention.

Explore More!

Research Paper