OncoEducate: A pilot study of generative AI to support patient education in GU cancer care.

Abstract

713 Background: Genitourinary (GU) cancer care involves complex treatment plans and goals, which can be difficult for patients to understand and navigate. Limited understanding can lead to distress and avoidable care utilization. We developed OncoEducate, a generative artificial intelligence (AI) tool that produces patient-friendly handouts, and piloted its accuracy, feasibility, and acceptability. Methods: OncoEducate uses the Claude Opus 4.1 large language model. User inputs include patient’s diagnosis, treatment intent, and treatment regimen. Based on these inputs, the tool generates tailored, two-page handouts summarizing diagnosis, treatment intent, regimen details, side effects, and reasons to contact the care team. Information is drawn from reputable public sources and standardized language is used for key concepts (e.g., treatment intent). Feedback from patient advocates, oncologists, and pharmacists informed iterative refinement. We conducted a prospective two-phase pilot study (IRB exempt, University of Pennsylvania). Phase 1: Nine handouts for common palliative-intent regimens in advanced kidney, prostate, and urothelial cancers were generated and reviewed by five GU oncologists and an advanced practice provider for accuracy, completeness, and readability. Feedback informed prompt refinement, yielding Version 2. Phase 2: Patients initiating these regimens received Version 2 handouts. Surveys assessed usefulness, readability, and acceptability (7-point Likert scale items) and comprehension of treatment intent (validated item from Cancer Care Outcomes Research and Surveillance study). Analyses were descriptive. Results: Clinicians rated Version 1 handouts as highly accurate and appropriate (median 6-7/7 across domains). Over eight weeks, 20 out of 21 eligible patients enrolled and completed surveys. Most were male (n = 17, 85%) with urothelial (n = 11, 55%) or kidney (n = 6, 30%) cancer. Median age was 71 (range: 46-83). Enfortumab vedotin + pembrolizumab (n = 8, 40%) was the most frequent regimen. Patients strongly agreed that the handouts were informative, easy to read, and improved their understanding of treatment plans and care team contact (median 7/7 for all). Comfort with AI-assisted education was high (median 7/7, range 2-7). 65% (n = 13) of patients correctly identified their treatment intent as palliative - exceeding historical benchmarks (19-31%, Weeks et al, NEJM 2012). Conclusions: OncoEducate handouts were well received by clinicians and patients and may improve understanding of treatment intent. Despite small sample size, findings demonstrate the feasibility of generative AI to deliver concise, personalized education. Larger randomized studies are needed to assess impact on patient outcomes.