Assistance of a large language model–based AI agent in preoperative communication for reduction of prostate cancer patients' anxiety and clinicians' burden: A prospective, randomized, single-blinded, phase II trial.

Abstract

56 Background: Preoperative communication is crucial in reducing patients' anxiety but demands considerable effort from clinicians. While large language models offer potential for addressing medical inquiries, their use in preoperative communication lacks sufficient clinical validation. Methods: We developed an AI agent using DeepSeek-R1 for preoperative communication, based on clinical guidelines and our institute's experience. Patients with newly diagnosed prostate cancer scheduled for radical surgery at our center were enrolled, excluding those with severe mental illness or cognitive dysfunction. Upon admission, patients were randomized at the ward level into control and AI agent-assisted groups. Both groups immediately received baseline emotional and cognitive assessments such as GAD-7 and B-IPQ. Specific questions of the two groups were collected via our AI agent. The AI agent-assisted group then received personalized responses from our AI agent, validated for scientific accuracy before delivery. Both groups then received preoperative communication from clinicians blinded to group assignments and underwent emotional and cognitive assessments again. Clinicians' workload was assessed using NASA-TLX. Primary endpoints were GAD-7 scores and NASA-TLX results. Secondary endpoints included other emotional (VAS-A, I-PANAS-SF), disease cognition scales (APAIS, B-IPQ), patient satisfaction, and communication time for clinicians. Sample size determination assumed 50% of patients would achieve a GAD-7 score below 4 with AI agent. With a two-sided significance level of 0.05 and power of 0.95, 100 patients per group are needed at least. Non-parametric rank-sum tests compared variables. All p values were two-sided (α = 0.05). Results: From Feb to Aug 2025, a total of 245 patients were randomized into control (n=124) and AI agent-assisted (n=121) groups. The two groups of patients totally asked 3,294 questions, averaging 13.94 questions per patient. These questions covered various aspects of surgery, including anesthesia, perioperative care, and management. Patients in the AI agent-assisted group experienced notably lower anxiety levels compared with the control group, with decrease in GAD-7 scores (median (IQR): 3.5 (2.0-5.0) vs. 7.5 (6.5-9.0), P<0.001). Clinician workload was notably reduced in the AI agent-assisted group, with lower NASA-TLX scores (median (IQR): 38.0 (24.5-52.5) vs. 57.0 (37.0-77.5), P<0.001) and remarkable nearly half the communication time compared to the control group (9.78 mins vs. 17.91 mins, P<0.001). Conclusions: AI-assisted preoperative communication significantly reduces patient anxiety, enhances disease perception, and decreases clinician workload, identifying previously unaddressed patient concerns. Clinical trial information: 07082049.