Extracorporeal Membrane Oxygenation (ECMO) Prediction of Mortality and Integrated Complication with Infection Risk Assessment by Machine Learning: Structured Narrative Review

Background: Extracorporeal membrane oxygenation (ECMO) is an extremely complex life-saving treatment that is performed on patients who are in severe cardiac and/or respiratory failure. Even with all the technological progress, ECMO is a procedure with high mortality and significant morbidity, making it important to predict the risks of ECMO in a timely and accurate manner. Objective: To offer a narrative review of the machine learning (ML) based risk prediction models currently used for ECMO, with particular focus on multi-outcome prediction—mortality, complications and infections. Methods: Structured narrative review of relevant literature was undertaken in PubMed, Web of Science, Scopus and IEEE Xplore from 2000 to August 2025. Eligible studies comprised machine-learning or artificial-intelligence algorithms to predict clinically important ECMO outcomes and in ECMO-supported patients. Studies that did not include conventional statistical methods, non-ECMO patient populations, or articles that presented editorials, commentaries, or articles lacking clinically relevant prediction outcomes were excluded. Titles and abstracts were screened; full text assessed for eligibility and extracted data organized using the CHARMS framework. Results: 25 studies were included, and these were divided into four areas, namely, mortality prediction, complication prediction (neurological, bleeding, thrombosis), requirement for ECMO and process management, and hospital readmission. In all these areas, ML models consistently showed better performance than traditional risk scores, picking up the high-dimensional interactions and nonlinear relationships. But the majority of current models are outcome-specific and use mostly static or pre-ECMO data. Of note, no potentially eligible ECMO-specific machine-learning model to predict incident secondary infection during ECMO support was found despite the impact of infection on outcomes, length of stay, antimicrobial usage, and health-care costs. Conclusions: The current evidence suggests that there is a shift in approach from traditional scoring methods to predictive approaches using ML. However, there are still key missing elements such as the lack of incorporation of dynamic physiological data, the lack of incorporation of immunological markers, and the lack of multi-outcome predictive approaches. These limitations need to be overcome to build clinically actionable decision-support systems. ML-based models are very promising in improving risk stratification in ECMO. Other studies should be conducted to create multi-outcome prediction models which incorporate mortality, complications, infection risk and real-time clinical data. These models could be integrated into routine ECMO practice, into electronic health records or ECMO monitoring systems, to generate time-updated and interpretable risk alerts to inform early intervention, antimicrobial stewardship, and resource planning.