Neural Network Model for Predicting Oxygen Transfer in Membrane Oxygenators Under Variable Extracorporeal Membrane Oxygenation Conditions

During extracorporeal membrane oxygenation (ECMO) management, early recognition and intervention are essential in cases of membrane oxygenator (MO) oxygenation failure. However, because the MO oxygen transfer (O2 transfer) capacity is influenced by various factors, clear evaluation criteria are lacking. Theoretical O2 transfer values provided by manufacturers are commonly used to assess MO performance; however, these values are presented only under standardized conditions. In this study, to develop an O2 transfer model for real-world ECMO settings, we conducted perfusion experiments using bovine blood under various venous blood compositions (hemoglobin concentration and oxygen saturation) and operational conditions (blood flow and fraction of delivered oxygen). Although substantial variability was observed in the relationships between O2 transfer and individual parameters, partial correlation analysis revealed significant associations with all factors, underscoring the need to incorporate them into the model. A multilayer feedforward neural network was employed to construct the model, achieving a high coefficient of determination (R2 = 0.992), demonstrating excellent predictive performance. The proposed O2 transfer model provides a framework for evaluating the oxygenation performance of MO under diverse ECMO conditions. By enabling comparison with real-time clinical data, it has the potential to support clinical decision-making and enhance the safety of ECMO management.