Can machine learning improve mortality prediction following cardiac surgery?

Area of Science:

• Cardiovascular Surgery
• Medical Informatics
• Machine Learning in Healthcare

Background:

• Presurgical risk assessment is critical for cardiac surgery patients due to high complication rates.
• Machine learning (ML) is increasingly explored for clinical risk prediction.
• Comparing ML algorithms with traditional models is essential for optimizing patient outcomes.

Purpose of the Study:

• To evaluate the performance of ML algorithms against logistic regression (LR) for predicting in-hospital mortality after cardiac surgery.
• To determine if ML offers superior predictive accuracy compared to established statistical methods.

Main Methods:

• A prospectively collected dataset of 28,761 adult cardiac surgery patients (1996-2017) was used.
• Models included neural network, random forest, naive Bayes, and retrained LR, using EuroSCORE features.
• Discrimination was assessed by AUC; calibration was evaluated using the calibration belt method.

Main Results:

• The in-hospital mortality rate was 2.7%.
• Retrained LR and random forest models demonstrated the best discrimination (AUC 0.80).
• All models exhibited significant miscalibration, with retrained LR showing the least calibration drift.

Conclusions:

• Machine learning methods did not show a significant advantage over logistic regression in predicting operative mortality.
• Logistic regression demonstrated comparable discrimination with better calibration and less drift.
• Further research may be needed to refine ML models for cardiac surgery risk prediction.