Reinforcement learning based automated anesthesia system for gastrointestinal endoscopy with a multicenter randomized trial.

Summary

In a multicenter RCT of 418 adults undergoing GI endoscopy, an RL-based automated system for ciprofol delivery achieved non-inferior hypoxemia rates to clinician-managed anesthesia and significantly shortened induction time without increasing drug use or recovery time. More intraoperative movement occurred under the automated system, consistent with lighter anesthetic depth.

Key Findings

• Hypoxemia incidence was similar between automated and clinician groups (14.42% vs 14.29%; OR 1.01, 95% CI 0.59–1.75; P=0.968).
• Induction time was shorter with automation (median 1.55 vs 1.90 minutes; P<0.001).
• No increase in total drug dose or recovery time; intraoperative body movement was more frequent under automated control.

Clinical Implications

Automated sedation for endoscopy could standardize safety, reduce induction time, and alleviate staffing pressures, with attention to movement management (e.g., adjuncts or depth targets).

Limitations

• Restricted to ASA I–II adults and endoscopy with ciprofol; generalizability to higher-risk populations and other agents is unknown
• Increased intraoperative movement suggests need for optimization of depth targets or adjuncts

Future Directions

Evaluate RL control across higher-risk cohorts, different procedures and agents, integrate movement/depth mitigation strategies, and assess workflow and cost-effectiveness.