Predictive Utility of a Machine Learning Algorithm in Estimating Mortality Risk in Cardiac Surgery.

Area of Science:

• Cardiovascular Surgery
• Medical Informatics
• Machine Learning

Background:

• Estimating operative mortality risk is crucial in cardiac surgery.
• Existing risk prediction models require continuous evaluation and improvement.

Purpose of the Study:

• To evaluate the predictive utility of an Extreme Gradient Boosting (XGBoost) machine learning algorithm.
• To compare the performance of XGBoost against the Society of Thoracic Surgeons Predicted Risk of Mortality (STS PROM) model.

Main Methods:

• Utilized data from 11,190 adult cardiac operations between 2011-2017.
• Developed and validated XGBoost models using 10-fold cross-validation and bootstrapping.
• Assessed model performance using precision, recall, calibration, C-index, accuracy, and F1 score.

Main Results:

• XGBoost showed moderate correlation with STS PROM overall (r=0.652) and weak correlation in patients with mortality (r=0.473).
• XGBoost outperformed STS PROM across multiple performance metrics, including C-index (0.808 vs 0.795) and F1 score (0.281 vs 0.230).
• XGBoost demonstrated improved accuracy and calibration compared to STS PROM.

Conclusions:

• Machine learning, specifically XGBoost, holds potential for enhancing predictive analytics in cardiac surgery.
• The observed modest improvements necessitate further validation in larger patient populations.