Abstract Aims To develop and evaluate machine learning–based models for predicting in-hospital mortality in post-cardiotomy cardiogenic shock patients supported with extracorporeal life support, aiming to improve prognostication and optimize clinical decision-making. Methods Data were obtained from the Extracorporeal Life Support Organization (ELSO) registry across 111 centres, including 5,982 adult patients who received extracorporeal life support for post-cardiotomy cardiogenic shock between January 2010 and December 2020. Data preprocessing comprised dataset integration, complete case analysis, and variable selection. Six machine learning algorithms, boosting, decision tree, k-nearest neighbours, random forest, naïve Bayes, and neural networks, were trained to predict in-hospital mortality and secondary clinical outcomes. The dataset was randomly divided into training (60%), validation (20%), and test (20%) cohorts. Results The boosting algorithm achieved the highest area under the curve (AUC = 0.759), followed by random forest (AUC = 0.688). Key predictors of in-hospital mortality included: age (survivors vs. non-survivors: 57.76 ± 14.63 vs. 61.25 ± 13.41 years, p 0.001), lactate during support (3.07 ± 2.88 vs. 5.79 ± 5.30 mmol/L, p 0.001), arterial pH (7.302 ± 0.122 vs. 7.278 ± 0.139, p 0.001), and BMI (29.12 ± 6.57 vs. 30.00 ± 7.12 kg/m², p 0.001). ECMO duration differed between groups (142.14 ± 150.91 vs. 150.50 ± 162.09 hours, p = 0.023); however, this on-support variable reflects clinical trajectory rather than a pre-initiation predictor and was excluded in a sensitivity analysis (random forest AUC = 0.697). Prediction of transplant-related outcomes was limited by class imbalance. Conclusions Machine learning models demonstrated moderate predictive performance for in-hospital mortality in post-cardiotomy cardiogenic shock patients on extracorporeal life support. The random forest model demonstrated moderate discriminative performance, highlighting the relevance of readily available clinical variables. External validation and calibration analysis are required before clinical implementation. The model is best interpreted as a tool for dynamic risk assessment during ECMO support.
Mahajna et al. (Sat,) studied this question.