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

Le résumé fourni par la source

The increasing demand for gastrointestinal endoscopic procedures, coupled with a global shortage of anesthesiologists, underscores the need for intelligent automation in anesthesia care. Reinforcement learning (RL) offers a promising strategy for autonomous anesthesia control, yet prospective clinical validation remains limited. We developed an RL-based automated anesthesia system for gastrointestinal endoscopy (AAS-GE) for automated ciprofol delivery and conducted a prospective, multicenter, randomized controlled trial across four centers in China between January 8 and August 27, 2025. Adults aged 18-65 years with American Society of Anesthesiologists physical status I-II undergoing gastrointestinal endoscopy were randomized 1:1 to receive either AAS-GE-controlled anesthesia or clinician-managed manual anesthesia. The primary outcome was the incidence of hypoxemia, defined as oxygen saturation below 92%, with secondary outcomes assessing hypoxemia severity, induction time, drug use, recovery, and adverse events. A total of 509 participants were included in algorithm development, and 418 were enrolled for clinical validation. The incidence of hypoxemia was comparable between groups (14.42 vs. 14.29%; odds ratio 1.01, 95% CI 0.59-1.75; P = 0.968), with no significant differences in secondary safety outcomes. AAS-GE achieved a shorter induction time (median 1.55 vs. 1.90 min; P < 0.001) without increasing total drug dose or recovery time. However, intraoperative body movement was more frequent in the AAS-GE group, consistent with lighter anesthesia depth. These results demonstrate the non-inferior safety and efficacy of AAS-GE compared with clinician management, supporting its potential to improve efficiency and standardize sedation care. Clinical registration: ClinicalTrials.gov on Feb. 26, 2025 (NCT06857344).

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.