Follow-up imaging recommendations vary across radiologists despite established guidelines. We evaluated whether a large language model (GPT-4o) can standardize follow-up timing and modality selection from routine radiology reports. In this retrospective, two-center study, 100 CT/MRI cases (25 each: head/neck, liver, lung, pancreas) were randomly sampled. GPT-4o and two human readers (R1, resident; R2, board-certified) generated follow-up recommendations from report text. Expert consensus, blinded to rater, assessed completeness, appropriateness of imaging modality, timing accuracy, and global follow up quality on a 5-point scale, with 5 being the highest score. Median global quality was 4 (2–5) for GPT-4o, 4 (1–5) for R1, and 4 (1–5) for R2; GPT-4o exceeded R1 (p < 0.01) and did not differ from 2 (p = 0.06). Relative treatment effect showed a reader effect (GPT4o 0.56 vs. R1 0.43 vs. R2 0.51; p < 0.01) and no center effect (0.50 vs. 0.50; p = 0.91). Correctness of follow-up timing was 96% (96/100) for GPT-4o, 75% (75/100) for R1, and 90% (90/100) for R2 (p < 0.001 for GPT-4o vs. R1). Completeness of follow-up was 92% (92/100) for GPT-4o, 91% (91/100) for R1, and 80% (80/100) for R2 (p = 0.014 for GPT-4o vs. Reader 2). No significant differences were observed regarding appropriateness of imaging modality. GPT-4o generated follow-up recommendations with overall quality comparable to an experienced radiologist and superior to a trainee, with high completeness and generally appropriate follow up timing and modality, supporting its role as decision support for standardized, guideline-aligned follow-up.
Kaya et al. (Thu,) studied this question.