Predictors of Mortality in Patients With ARDS Undergoing Extracorporeal Membrane Oxygenation

Tuesday, March 24, 2026

11:15 AM - 12:15 PM Central Time

Location: Connections Central - RST 01

Maximiliano Kakazu, MD

Corewell Health - Michigan State University, Michigan

Disclosure(s): InflaRx Pharmaceuticals: Advisory Board (Terminated, May 17, 2024)
SA

Shrestha Acharya, N/A

Data Scientist
University of Colorado

Disclosure information not submitted.
AR

Asad Rehman, N/A

Neurointensivist
Northwest Community Healthcare

Disclosure information not submitted.
RA

Rafi Al Attrach, N/A

Data Scientist
MIT

Disclosure information not submitted.
KE

Khalil El Gharib, N/A

Pulmonary and Critical Care Fellow
Rutgers Robert Wood Johnson Medical School

Disclosure information not submitted.
RT

Riley Thompson, N/A

Data Scientist
Mayo Clinic

Disclosure information not submitted.
HB

Han Byul Lee, n/a

Data Scientist
University of Southern California

Disclosure information not submitted.

Introduction: Venous- venous Extracorporeal Membrane Oxygenation (VV-ECMO) is indicated in severe ARDS when conventional management fails. It is a limited resource and triaging of patients is important. This study aims to evaluate predictors of mortality in patients with ARDS undergoing VV-ECMO.

Methods: We performed a retrospective study using the REDISCOVER-ICU dataset derived from electronic health records in 7 centers. Adults with diagnosis of acute respiratory failure who underwent VV-ECMO were evaluated. LASSO analysis was used for feature selection. SHAP analysis was used for explainability. Random Forest was used for missing imputations. Random Forest, XG Boost Model, Support Vector Machine (SVM) and Multilayer Perceptron were used to build a predictive model for mortality. MIMIC-IV was used for validation of the predictive model.

Results: 175 patients were included in the cohort. Eight variables were selected with LASSO feature selection: average PaO2, PaO2/FiO2 ratio, average diastolic blood pressure, average platelets, ethnicity, average ferritin, race, and BMI. Not all variables were available for all the patients. Receiver Operating Characteristic for SVM Model was 0.76, Random Forest 0.77, XG Boost Model 0.56, and MLP Model was 0.60. The confusion matrix for the Random Forest Model reported an accuracy of 65%, precision of 73% and sensitivity of 70%. This model was validated in MIMIC-IV. Out of 8 predictive factors, 4 are most clinically relevant, oxygen levels, PaO2/FIO2 ratio, diastolic blood pressure and platelet count.

Conclusions: Based on our analysis of REDISCOVERY-ICU, eight variables were predictors of mortality in patients with ARDS undergoing VV-ECMO. Random Forest Model performed the best to predict mortality using these factors. The model was validated in MIMIC-IV but unbalanced data limits the generalizability of this model.