Artificial intelligence (AI) based prediction models for in hospital mortality in patients undergoing extracorporeal life support for post cardiotomy shock

Objective: This study focuses on developing a machine learning driven predictive models for clinical outcomes in post cardiotomy patients receiving extracorporeal membrane oxygenation (ECMO). The aim is to enhance prognostication and resource allocation within critical care settings.
Design: We utilized a comprehensive dataset from the Extracorporeal Life Support Organization (ELSO) registry, comprising 5,882 patients who received veno arterial ECMO between January 1, 2010, and December 31, 2020. Six machine learning algorithms (boosting, decision tree, k-nearest neighbors, random forest, naive bayes, and neural network) were applied to predict key outcomes e.g died or poor prognosis, in hospital mortality, and transplant etc.
Setting: There were data preprocessing like merging clinical datasets, handling missing data, and selecting relevant variables. The dataset was split into 70% training and 30% testing sets for model evaluation.
Results: The random forest model outperformed other algorithms in predicting mortality related outcomes. Key predictors identified were lactate levels, ECMO duration, and age etc. Elevated lactate levels during ECMO were strongly associated with poor prognosis and higher in hospital mortality. Prolonged ECMO support and older age were critical factors for predicting both poor prognosis and mortality. The model faced challenges in predicting transplant outcomes due to the imbalanced dataset.
Conclusion: This research demonstrates the potential of machine learning in improving patient care and outcomes in post cardiotomy ECMO patients. The random forest model showed robust performance in predicting mortality related outcomes, highlighting the importance of key clinical variables. Future work should focus on refining the model with new data and advanced algorithms, integrating it into clinical practice for real time prognostication, and expanding the model to predict other clinically relevant outcomes.