Conversational AI Establishes Patient Understanding Before Clinician Contact: A Randomized Equivalence Trial (Preprint)

BACKGROUND Informed consent depends on patients' understanding of procedural risks, yet comprehension of anesthesia risk remains poor despite routine preoperative consultation. Current consent processes rely on time-limited clinician encounters as the primary mechanism of patient education, creating challenges for scalability, efficiency, and patient engagement. Conversational artificial intelligence (AI) offers an alternative model in which foundational education occurs before clinician contact, potentially redefining the role of the consultation itself. OBJECTIVE We evaluated whether a multilingual retrieval-augmented conversational AI system could achieve patient-reported understanding comparable to standard consultation while improving clinical workflow efficiency. METHODS We conducted a prospective randomized equivalence trial involving 130 adults undergoing elective surgery at a tertiary academic medical center. Participants were randomly assigned in a 1:1 ratio to receive either PEAR (Preoperative Education of Anaesthesia Risks), a multilingual retrieval-augmented conversational AI system grounded in institutional consent materials, followed by standard consultation, or standard consultation alone. The primary outcome was patient-reported understanding of anesthesia risk after consultation, assessed using three 5-point Likert-scale measures evaluating understanding of risks, confidence in the anesthesia plan, and ability to recall and explain key risks. Equivalence was defined a priori as a margin of ±0.5 Likert points. Secondary outcomes included pre-consultation understanding, technology acceptance, patient preference, clinical efficiency, safety, and economic impact. RESULTS Among 130 enrolled participants (mean age, 52.4 years; 54.6% male), post-consultation understanding in the AI group met the prespecified equivalence criterion across all primary measures. Mean between-group differences (AI minus control) were −0.03 (90% confidence interval [CI], −0.19 to 0.13), −0.05 (90% CI, −0.22 to 0.13), and 0.12 (90% CI, −0.07 to 0.32), respectively; all confidence intervals were contained within the equivalence margin. Notably, understanding scores obtained immediately after AI interaction and before clinician contact were not significantly different from post-consultation scores in the control group, suggesting that patient understanding had been established before the clinical encounter. Technology acceptance was high across all domains. Sixty-three percent of participants preferred AI-assisted education to traditional consultation. PEAR reduced combined consultation and documentation time by 19.3 minutes per patient, corresponding to an estimated annual net benefit of SGD 0.99 million (USD 0.78 million) at a single tertiary hospital. Clinicians identified documentation inaccuracies in 16.9% of encounters, all minor or moderate in severity, supporting continued clinician oversight. CONCLUSIONS A retrieval-augmented conversational AI system achieved patient-reported understanding of anesthesia risk equivalent to standard consultation while substantially improving workflow efficiency. These findings suggest that patient education can be shifted upstream of clinician encounters, enabling consultations to focus on verification, contextualization, and shared decision-making rather than primary information delivery. CLINICALTRIAL The study was approved by the SingHealth Centralised Institutional Review Board (CIRB 2025/0673), registered at ClinicalTrials.gov (NCT06949462), and reported in accordance with CONSORT 2025 and CONSORT-AI Extension guidelines.