JMIR Publications

ABSTRACT

Background:

AI-generated instructional images are increasingly used in rehabilitation patient education, yet single-pass generation yields anatomical inaccuracies and unsafe postures.

Objective:

Whether structured multi-agent AI critique pipelines measurably improve clinician-judged image quality has not been empirically evaluated.

Methods:

In this double-blind comparative study, 41 reviewers from three independent hospital and university teams rated 98 paired rehabilitation exercise images on a validated 14-item instrument covering four domains: clinical accuracy, instructional utility, patient safety, and clinical adoption intent. Image A was produced by single-pass AI generation; Image B underwent a four-agent critique-and-refinement pipeline. Wilcoxon signed-rank tests with Bonferroni correction, mixed-effects models, and intraclass correlation coefficients were applied.

Results:

Results Across 815 paired ratings, all 14 rating dimensions favored Image B after Bonferroni correction (all p < 0.001; Rank-Biserial Correlation, r_RBC 0.48–0.73), with the largest effects in visual clarity (r_RBC 0.71–0.73) and patient safety (r_RBC 0.65–0.70). At construct level, patient safety showed the greatest absolute improvement, followed by instructional utility, clinical adoption intent, and clinical accuracy. Image B was the preferred choice in 65.3% of forced-choice judgements versus 12.4% for Image A (one-sided binomial p < 0.001). Open-text defect coding showed Image B had markedly lower rates of unclear imagery (9.7% vs. 2.6%), lack of instructions (6.7% vs. 0.1%), and missing safety notes (2.5% vs. 0.4%). Quality advantages were observed across all ten anatomical regions and were independent of professional backgrounds.

Conclusions:

A multiple-agent critique-and-refinement pipeline substantially improves clinician-judged quality of AI-generated rehabilitation instructional images across all measured domains, with the greatest gains in patient safety and visual clarity. These findings provide the first empirical quality benchmark for pipeline-based AI image generation in rehabilitation and support multi-AI review as a scalable minimum standard before clinical deployment, pending prospective patient-outcome evaluation.