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

<h2>Abstract</h2><h3>Objective</h3> Prolonged mechanical ventilation after cardiac surgery contributes significantly to morbidity, mortality, and excessive hospital resource use. Accurate prediction of prolonged mechanical ventilation duration can improve decision-making and patient outcomes. We aimed to develop and validate time-to-event models to predict the duration of ventilation and prolonged mechanical ventilation. <h3>Methods</h3> From the Medical Information Mart for Intensive Care III and IV databases, we extracted postoperative data from all cardiac surgery patients. We benchmarked 3 machine learning time-to-event algorithms (random survival forest, gradient boosted survival model, and survival support vector machine) against traditional Elastic-Net Cox regression. We evaluated model performance using weighted mean area under the curve (<mml:math><mml:mrow><mml:mrow><mml:mi>A</mml:mi><mml:mi>U</mml:mi><mml:mi>C</mml:mi></mml:mrow><mml:mo>¯</mml:mo><mml:mi>w</mml:mi><mml:mrow><mml:mi>C</mml:mi><mml:mo>,</mml:mo><mml:mi>D</mml:mi></mml:mrow></mml:mrow></mml:math>), cumulative/dynamic area under the receiver operating characteristic curve (AUC^C,D(<i>t</i>)), Concordance Index, and integrated Brier score. Permutation feature importance was reported for the best models. We conducted a sensitivity analysis to evaluate model fairness across different races and sexes. <h3>Results</h3> Models were trained on data from 10,430 cardiac surgery patients ventilated for a median of 7.0 hours (interquartile range, 4.4-16.0). Random survival forest had the highest <mml:math><mml:mrow><mml:mrow><mml:mi>A</mml:mi><mml:mi>U</mml:mi><mml:mi>C</mml:mi></mml:mrow><mml:mo>¯</mml:mo><mml:mi>w</mml:mi><mml:mrow><mml:mi>C</mml:mi><mml:mo>,</mml:mo><mml:mi>D</mml:mi></mml:mrow></mml:mrow></mml:math> (0.834, 95% CI, 0.832-0.836) and integrated Brier score (0.041), whereas gradient boosted survival model had the highest Concordance Index (0.721, 95% CI, 0.717-0.724). All machine learning models significantly outperformed Elastic-Net Cox Regression. Ventilatory settings, laboratory results, and Sequential Organ Failure Assessment score within 4 hours of intubation were identified as the most important features. Sensitivity analysis showed equal or improved performance for minority female and non-White cohorts. <h3>Conclusions</h3> Machine learning time-to-event models for prolonged mechanical ventilation and the duration of ventilation, particularly random survival forest and gradient boosted survival model, have significantly improved performance compared with current state-of-the-art tools and may be valuable decision supports in the postoperative management of cardiac surgery patients.