Machine learning time-to-event algorithms for predicting the duration of ventilation after cardiac surgery

Area of Science:

• Cardiology
• Critical Care Medicine
• Data Science & Machine Learning

Background:

• Prolonged mechanical ventilation (PMV) post-cardiac surgery is linked to increased morbidity, mortality, and healthcare costs.
• Predicting PMV duration is crucial for optimizing patient care and resource allocation.

Purpose of the Study:

• To develop and validate time-to-event models for predicting ventilation duration and PMV in cardiac surgery patients.
• To compare the performance of machine learning algorithms against traditional regression models.

Main Methods:

• Utilized data from the Medical Information Mart for Intensive Care III and IV databases for 10,430 cardiac surgery patients.
• Benchmarked random survival forest, gradient boosted survival model, and survival support vector machine against Elastic-Net Cox regression.
• Evaluated models using weighted mean AUC, Concordance Index, and integrated Brier score, with sensitivity analysis for fairness.

Main Results:

• Machine learning models, particularly random survival forest and gradient boosted survival model, significantly outperformed Elastic-Net Cox regression.
• Random survival forest achieved the highest weighted mean AUC (0.834) and best integrated Brier score (0.041).
• Key predictors included ventilatory settings, lab results, and Sequential Organ Failure Assessment score within 4 hours of intubation.

Conclusions:

• Machine learning time-to-event models demonstrate superior performance in predicting ventilation duration and PMV.
• These models offer valuable decision support for postoperative management in cardiac surgery.
• Fairness analysis indicated comparable or improved performance across different racial and sex groups.