Predicting Short-Term Mortality in ECMO-Supported Patients Secondary to Decompensated Heart Failure and Acute Myocardial Infarction Using Machine Learning

Abstract

<h3>Purpose</h3> Patients with refractory cardiogenic shock (Stage E) face imminent death and often require extra-corporeal membrane oxygenation (ECMO) to achieve stabilization and survival. ECMO is a resource intense therapy and with high morbidity and mortality. The purpose of our study was to develop a prediction model, using machine learning, of short-term mortality in patients with decompensated heart failure (DHF) and acute myocardial infarction (AMI) who required ECMO. <h3>Methods</h3> Patients supported by VA-ECMO due to DHF or AMI registered in the Spectrum Health ECMO registry were included in this study. Clinical, echocardiographic, laboratory and hemodynamic characteristics were obtained in all patients. Thirty-day survival from ECMO cannulation was calculated using Kaplan-Meier methodology. Using machine learning (elastic-net method) a predictive model for 30-day mortality was developed in the derivation cohort, the model was then tested on the validation cohort. <h3>Results</h3> A total of 283 patients met the inclusion criteria (228 DHF and 55 AMI). Of these, 151 died within the first 30 days post ECMO insertion. Survivors had similar characteristics to non-survivors except for lower troponin and creatinine levels. Machine learning identified 8 variables associated with 30-day mortality in the derivation cohort (Troponin, eGFR, total bilirubin, BMI, lactate, ethnicity, age and pH) with AUC of 0.68 (figure 1 A). The importance of each variable in the predicting model is shown in Figure 1B. Our score predicted 30 days mortality in the validation cohort with a good accuracy (AUC: 0.72, Figure 1C). <h3>Conclusion</h3> Using the machine learning algorithm, we developed a model for predicting 30-day survival for cardiogenic shock patients supported by ECMO due to DHF and AMI.