A Prospective Controlled Trial of Large Language Model–based Simplification of Oncologic CT Reports for Patients with Cancer

Abstract

Background Radiology staging reports (ie, oncologic reports) are written for referring physicians using complex medical terminology. Large language models (LLMs) show promise for simplifying medical text for patient use, but controlled studies evaluating the impact of LLM simplification on patients' comprehension of radiology reports are lacking. Purpose To evaluate whether LLM-based simplification of oncologic CT reports improves patients' cognitive workload, text comprehension, perception, and reading time. Materials and Methods This prospective, controlled, open-label, quasi-randomized trial enrolled 200 adults with cancer who underwent routine CT restaging. Between April and May 2025, participants were alternately assigned to receive either standard CT reports (100 participants) or LLM-simplified versions created using Llama 3.3 70B (Meta) with mandatory radiologist review (100 participants). The primary outcomes were participant-reported scores on nine seven-point Likert scale items, and composite scores, in the domains of cognitive workload, text comprehension, and report perception, as well as reading time. Secondary outcomes included readability metrics and independent radiologist assessments of report errors, usefulness, and quality. Statistical analyses included logistic regression adjusted for participant characteristics. Results Among the 200 participants (mean age, 64 years ± 14 [SD]; 112 male participants), simplified reports reduced the median reading time from 7 minutes to 2 minutes (<i>P</i> < .001). Participants who received simplified reports reported lower cognitive workload (adjusted odds ratio [OR], 0.18 [95% CI: 0.13, 0.25]), better comprehension (adjusted OR, 13.28 [95% CI: 9.31, 18.93]), and better perception of report usefulness (adjusted OR, 5.46 [95% CI: 3.55, 8.38]) than did those who received standard reports (all <i>P</i> < .001). Simplification improved report readability (mean Flesch-Kincaid Grade Level, 8.89 ± 0.93 vs 13.69 ± 1.13; <i>P</i> < .001). Radiologist review revealed factual errors in 6% (moderate, 2%; severe, 4%), content omissions in 7% (minor, 2%; moderate, 1%; severe, 4%), and inappropriate additions in 3% (minor, 1%; moderate, 2%) of simplified reports. Conclusion LLM simplification of oncologic CT reports improved patient comprehension and reduced reading burden. However, clinically relevant errors were identified. © RSNA, 2025 <i>Supplemental material is available for this article.</i>

Ähnliche Arbeiten