AI-assisted preoperative anesthesia evaluation workflow versus conventional bedside evaluation: a randomized controlled non-inferiority trial

Abstract Background Digital tools may improve the efficiency and standardization of preoperative anesthesia assessment, but evidence comparing AI-assisted workflows with conventional bedside evaluation remains limited. We compared ASA physical status classification concordance, documentation quality, recorded assessment duration, and patient-reported experience between an AI-assisted clinical workflow and conventional bedside evaluation. Methods In this single-center, open-label randomized controlled non-inferiority trial at Huashan Hospital, 600 hospitalized adults aged 18–65 years scheduled for elective non-cardiac surgery under general anesthesia were allocated 1:1 to conventional bedside evaluation or an AI-assisted workflow. In the intervention group, patients or caregivers completed a deterministic rule-based assessment using a WeChat mini-program, after which an anesthesiology resident reviewed and confirmed the report. The primary outcome was exact concordance between the initial workflow ASA classification and the independent day-of-surgery classification assigned by an attending anesthesiologist who was unaware of allocation and the initial classification. The stated non-inferiority margin was − 5 percentage points, with a one-sided alpha of 0.025. All randomized participants were included in both intention-to-treat (ITT) and per-protocol (PP) analyses. Results ASA concordance was 93.3% (280/300) in the AI-assisted group and 91.7% (275/300) in the conventional group. The risk difference was + 1.67 percentage points (95% CI − 2.55 to + 5.88). Relative to the − 5-percentage-point margin, Z = 3.10 and the one-sided P value was 0.001; ITT and PP results were identical. Grade A documentation occurred in 197/300 (65.7%) AI-assisted reports and 95/300 (31.7%) conventional reports. Median recorded assessment time was 2.0 min (IQR 2.0–3.0) versus 6.0 min (IQR 4.0–7.0), with a median difference of − 4.0 min (bootstrap 95% CI − 4.0 to − 2.5; P < 0.001). Overall satisfaction (agree/strongly agree) was reported by 262/300 (87.3%) and 234/300 (78.0%) participants, respectively (risk difference + 9.33 percentage points, 95% CI + 3.32 to + 15.34; continuity-corrected P = 0.003). Full ordinal response distributions did not differ for procedure understanding, concern resolution, or anxiety; understanding of anesthesia-related risks was less favorable in the AI-assisted group ( P = 0.001). Conclusions Among selected adults aged 18–65 years at one institution, the AI-assisted workflow met the stated non-inferiority criterion for ASA concordance, shortened recorded assessment time, and produced more complete documentation. Higher overall satisfaction did not correspond to improved communication outcomes. The findings do not establish clinical safety, cost-effectiveness, or generalizability to older, higher-risk, emergency, or digitally disadvantaged populations. Trial registration Chinese Clinical Trial Registry, ChiCTR2400086869; registered 12 July 2024, after participant enrollment had ended.