Advancing Cardiac Care with Machine Learning: Predicting Major Outcomes
In the evolving landscape of healthcare, machine learning (ML) is emerging as a promising tool to enhance predictive capabilities and improve patient outcomes. A recent initiative aimed to harness ML’s potential by developing a model that predicts the risk of major cardiac events in patients admitted to the cardiology service. This article delves into the objectives, methods, and results of this groundbreaking study.
Objectives
The primary goal was to create a robust ML model using electronic health record data to forecast the likelihood of a significant cardiac event within three months. The events of interest included admission to a ventricular assist device, placement on a heart transplant waiting list, or death. The study also sought to evaluate the model’s accuracy in a prospective silent trial, where predictions were not disclosed to physicians.
Methods and Analysis
Data for the study were meticulously gathered from enrollments in the cardiology service over two distinct periods: June 2, 2018, to August 21, 2023, for retrospective analysis, and May 10, 2024, to October 26, 2024, for prospective evaluation. The data source, SickKids Enterprise-wide Data in the Azure Repository, is renowned for its reliability and thorough validation. Predictions were made based on data recorded the previous day.
The study employed advanced ML techniques, including L2 regularized logistic regression, LightGBM, and XGBoost. The training cohorts consisted of the target cohort as well as all inpatient admissions, aiming to ensure a comprehensive and representative model.
Results
The LightGBM model emerged as the best performer during the retrospective phase, having been trained across all inpatient data. The retrospective phase involved analyzing 51,571 approvals, while the prospective silent trial included 515 approvals. The final model incorporated a substantial 7,553 features, underscoring the complexity and depth of the analysis.
Success was measured by the area under the receiver operating characteristic curve (AUC), which was 0.88 (95% CI 0.88 to 0.89) during the retrospective phase and 0.82 (95% CI 0.79 to 0.83) in the prospective silent trial. The silent study also demonstrated a positive predictive value of 0.19 and a notably high negative predictive value of 0.97, based on thresholds determined in the retrospective phase.
Conclusions
This study marks a significant step forward in predicting major cardiac outcomes through machine learning. By integrating real-world data into a mission-oriented framework, the research highlights ML’s potential in augmenting clinical decision-making. Future goals include post-deployment evaluations to refine and validate the model further.
For more details, please refer to the source Here.
“`
