rECMOmender: Reinforcement Learning for Decision Support in Venovenous Extracorporeal Membrane Oxygenation Management
Year: 2026
Type: article
Abstract: CONTEXT: Management of ventilator and venovenous extracorporeal membrane oxygenation (ECMO) settings in critically ill adults requires individualized decisions to balance oxygenation, ventilation, and complication risks. Existing approaches rely heavily on clinician experience, with limited decision support tools available. HYPOTHESIS: An offline reinforcement learning agent trained on real-world venovenous ECMO clinical data can generate safe, interpretable, and clinically aligned recommendations for ECMO and ventilator management, including support for earlier and more efficient weaning. METHODS AND MODELS: We conducted a retrospective study using electronic health record data from 184 adult patients who underwent venovenous ECMO at a tertiary care center. rECMOmender was developed using conservative Q-learning with a physiologically informed reward structure. Multiple model variants were compared across discrete and continuous action spaces and two reward formulations. Performance was assessed using fitted Q evaluation, comparison of action distributions, and alignment with clinician practice. RESULTS: rECMOmender generated stable, interpretable recommendations across five key parameters: Fio2, positive end-expiratory pressure (PEEP), respiratory rate, sweep gas flow, and blood flow rate. It selected Fio2 values in the 40-50% range most frequently (46.75% vs. 45.65% for clinicians) and favored PEEP of 9-11 cm H2O (43.94% vs. 34.28%), while using high PEEP settings (13-20 cm H2O) 73.43% less often. Compared with clinicians, rECMOmender increased large parameter shifts (> 1 bin) by 72.53% for Fio2, 348.15% for PEEP, 299.21% for respiratory rate, 96.68% for sweep gas, and 34.16% for blood flow, resulting in an overall 120.37% increase in major adjustments (3105 vs. 1409). INTERPRETATIONS AND CONCLUSIONS: rECMOmender demonstrated dynamic but safety conscious adjustments that aligned with clinical patterns, indicating potential as a decision support tool that complements clinician judgment.
Source: Critical Care Explorations
Language: English
Subfield: Biomedical Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $1,600
Awards: Scalable and Interoperable framework for a clinically diverse and generalizable sepsis Biorepository using Electronic alerts for Recruitment driven by Artificial Intelligence (short title: SIBER-AI), Duke-UNC Collaborative Pediatric Clinical Pharmacology Postdoctoral Training Program, Unified Program for Therapeutics in Children, Sepsis Physiomarkers for Appropriate Risk Knowledge of monitored patients in the ICU (SPARK-ICU)
OpenAlex ID: W7128164351