A High Granularity Machine Learning Model to Predict Successful Weaning During V-V ECMO

Weaning from venovenous extracorporeal membrane oxygenation (V-V ECMO) remains unstandardized due to limited literature, with delayed weaning increasing the risk of vascular injury and multi-organ failure. We aim to assess whether high-granularity data, involving longitudinal ECMO, physiological, respiratory, and hemodynamic electronic medical record (EMR) variables, can predict V-V ECMO weaning outcomes and outperform the Respiratory ECMO Survival Prediction (RESP) score.

We conducted a retrospective review of adult V-V ECMO patients (≥ 18 years) at Johns Hopkins Hospital from 2016 to 2024. Machine learning (ML) models were developed using the RESP score alone (Model A) and EMR time-series data across four intervals: pre-ECMO, first 24 h, last 24 h before decannulation, and full ECMO run (Models B-E). Random Forest, CatBoost, AdaBoost, and XGBoost algorithms were trained using 80/20 train-test splits. Feature importance was assessed using Shapley Additive Explanations (SHAP) values.

119 V-V ECMO patients (median age = 48%, 59.6% female) were included, with 62 (52.1%) experiencing mortality at discharge. Model E attained the highest performance with an area under the receiver operating characteristic curve (AUROC) of 0.904; Models A-D achieved AUROCs of 0.815, 0.715, 0.515, and 0.881 respectively. SHAP analysis identified Body Mass Index (BMI), pulse, SpO₂, and immunocompromised status as key features in predicting weaning success. BMI and pulse were strongest in the first 24 h, whereas age was most influential pre-ECMO.

High-granularity ML models accurately predicted V-V ECMO weaning, with BMI, age, and immunocompromised status as key predictors. Their superior performance against RESP score-based ML models suggests clinical utility for V-V ECMO weaning.