Comparing Artificial Intelligence–Generated and Clinician-Created Personalized Self-Management Guidance for Patients With Knee Osteoarthritis: Blinded Observational Study

Background Knee osteoarthritis is a prevalent, chronic musculoskeletal disorder that impairs mobility and quality of life. Personalized patient education aims to improve self-management and adherence; yet, its delivery is often limited by time constraints, clinician workload, and the heterogeneity of patient needs. Recent advances in large language models offer potential solutions. GPT-4 (OpenAI), distinguished by its long-context reasoning and adoption in clinical artificial intelligence research, emerged as a leading candidate for personalized health communication. However, its application in generating condition-specific educational guidance remains underexplored, and concerns about misinformation, personalization limits, and ethical oversight remain. Objective We evaluated GPT-4’s ability to generate individualized self-management guidance for patients with knee osteoarthritis in comparison with clinician-created content. Methods This 2-phase, double-blind, observational study used data from 50 patients previously enrolled in a registered randomized trial. In phase 1, 2 orthopedic clinicians each generated personalized education materials for 25 patient profiles using anonymized clinical data, including history, symptoms, and lifestyle. In phase 2, the same datasets were processed by GPT-4 using standardized prompts. All content was anonymized and evaluated by 2 independent, blinded clinical experts using validated scoring systems. Evaluation criteria included efficiency, readability (Flesch-Kincaid, Gunning Fog, Coleman-Liau, and Simple Measure of Gobbledygook), accuracy, personalization, and comprehensiveness and safety. Disagreements between reviewers were resolved through consensus or third-party adjudication. Results GPT-4 outperformed clinicians in content generation speed (530.03 vs 37.29 words per min, P<.001). Readability was better on the Flesch-Kincaid (mean 11.56, SD 1.08 vs mean 12.67 SD 0.95), Gunning Fog (mean 12.47, SD 1.36 vs mean 14.56, SD 0.93), and Simple Measure of Gobbledygook (mean 13.33, SD 1.00 vs mean 13.81 SD 0.69) indices (all P<.001), though GPT-4 scored slightly higher on the Coleman-Liau Index (mean 15.90, SD 1.03 vs mean 15.15, SD 0.91). GPT-4 also outperformed clinicians in accuracy (mean 5.31, SD 1.73 vs mean 4.76, SD 1.10; P=.05, personalization (mean 54.32, SD 6.21 vs mean 33.20, SD 5.40; P<.001), comprehensiveness (mean 51.74, SD 6.47 vs mean 35.26, SD 6.66; P<.001), and safety (median 61, IQR 58-66 vs median 50, IQR 47-55.25; P<.001). Conclusions GPT-4 could generate personalized self-management guidance for knee osteoarthritis with greater efficiency, accuracy, personalization, comprehensiveness, and safety than clinician-generated content, as assessed using standardized, guideline-aligned evaluation frameworks. These findings underscore the potential of large language models to support scalable, high-quality patient education in chronic disease management. The observed lexical complexity suggests the need to refine outputs for populations with limited health literacy. As an exploratory, single-center study, these results warrant confirmation in larger, multicenter cohorts with diverse demographic profiles. Future implementation should be guided by ethical and operational safeguards, including data privacy, transparency, and the delineation of clinical responsibility. Hybrid models integrating artificial intelligence–generated content with clinician oversight may offer a pragmatic path forward.