Performance of a Machine Learning Algorithm in Predicting Outcomes of Aortic Valve Replacement

Area of Science:

• Cardiovascular Surgery
• Medical Informatics
• Machine Learning

Background:

• Surgical aortic valve replacement (SAVR) is a critical procedure with established risk prediction models.
• Evaluating novel methods for improving outcome prediction in SAVR is essential for patient care.

Purpose of the Study:

• To assess the performance of an extreme gradient boosting (XGBoost) machine learning algorithm for predicting outcomes in isolated SAVR.
• To compare the predictive accuracy of XGBoost against existing Society of Thoracic Surgeons (STS) models.

Main Methods:

• Utilized a large dataset of 243,142 adult patients undergoing isolated SAVR from the STS National Database (2007-2017).
• Randomly split data into training (4:1) and validation sets for model development and testing.
• Employed XGBoost, evaluating model calibration (observed-to-expected ratio, calibration-in-the-large, slope) and discrimination (c-index).

Main Results:

• XGBoost exhibited excellent calibration across multiple metrics.
• The algorithm demonstrated significantly improved discrimination (c-index) compared to STS models for 5 of 7 outcomes, including operative mortality and acute renal failure.
• Comparable predictive performance was observed for stroke and deep sternal wound infection.

Conclusions:

• The XGBoost machine learning algorithm shows excellent calibration and improved discriminatory ability for predicting outcomes in isolated SAVR.
• These findings suggest that ML algorithms can enhance the precision of risk stratification for SAVR patients.
• Further integration of ML tools could refine clinical decision-making and patient management in cardiac surgery.