Use of Artificial Intelligence to Predict Outcomes in Critically Ill Patients Receiving Extracorporeal Membrane Oxygenation: A Systematic Review.

Extracorporeal membrane oxygenation (ECMO) is widely used in patients with severe cardiac or respiratory failure. ECMO is resource and cost intensive and carries the potential for significant complications; therefore, identifying patients who would derive the most benefit and personalizing its application are important. A number of scoring systems have previously been developed to predict the outcomes of patients receiving ECMO. Machine learning (ML)-based models are highly effective at incorporating the large amount of complex data and large number of variables that may be associated with outcomes. The authors conducted a systematic review to identify and critically appraise ML-based models developed to predict ECMO outcomes, including mortality, mortality surrogates, and complications, with particular attention to their methodologic rigor and potential for clinical implementation. In this systematic review, 5 databases (MEDLINE, Web of Science, Scopus, Cumulative Index to Nursing and Allied Health Literature [CINAHL], and Embase) were searched for eligible published studies. Each database was searched from inception to the date of search. Fourteen articles were identified, reporting the results of 31 different models. The most common outcomes predicted were mortality (n = 6) and neurologic outcomes (n = 3). A variety of other clinically important endpoints were used, including ECMO complications (n = 3), successful liberation from ECMO (n = 1), and readmission to hospital (n = 1). The most used ML models were random forest (n = 10) and extreme gradient boosting (n = 7). Three models (7.3%) were validated in an external population. The area under the receiver operating characteristic curve values ranged from 0.33 to 0.942. Twenty models (49%) demonstrated good to excellent predictive performance; however, all included models were judged to be at high risk of bias primarily because of the model analysis and model predictors. ML-based models hold the potential to be used as decision-support tools for patients receiving ECMO. However, their clinical applicability is limited by variability in discriminatory performance, consistently high risk of bias, and infrequent application of external validation.