Building an Extracorporeal Membrane Oxygenation Digital Twin Using High-Resolution Patient Data: An artificial intelligence model for virtual reality simulation

OBJECTIVES: Extracorporeal membrane oxygenation (ECMO) is a life-saving therapy for severe cardiopulmonary failure, but structured training remains constrained by costs, logistics, and the absence of validated high-fidelity simulators. This study aimed to develop an ECMO digital twin capable of supporting training in virtual reality (VR). METHODS: We integrated high-frequency ECMO machine data with electronic health record information from 335 patients across two centres. Data streams were synchronised at a 30-second resolution. A hierarchical two-stage system was designed: Model 1 predicted ECMO device outputs, while Model 2 combined those outputs with patient vital functions such as heart rate and blood pressure. This model was integrated into a VR simulation and underwent testing by 21 experts. RESULTS: Model 2 demonstrated Root Mean Square Errors (RMSE) of 15.23 mmHg (diastolic arterial blood pressure), 19.50BPM (heart rate), 2.94% (peripheral oxygen saturation), and 1.42 mmHg (end-tidal carbon dioxide) on the test set. Neural networks produced clinically coherent predictions. The models were implemented in an Unreal Engine-based VR simulator using the open neural network exchange format, with real-time latency inference and scenario switching. Expert testing confirmed good performance and clinically plausible physiological responses in the simulation. CONCLUSIONS: High-resolution ECMO data can be transformed into a digital twin for VR training. This framework broadens access to advanced ECMO education and establishes a foundation for multicentre validation, federated learning, and future expansion towards a critical-care digital-twin platform.