Enhanced PROcedural Information TRANSLATability for Patient-Centered Care in Interventional Radiology with Large Language Models (PRO-TRANSLATE IR)

Purpose: To utilize GPT-4 to generate multilingual, patient facing interventional radiology instructions and evaluate their clinical appropriateness and lay comprehension across eight languages.

Materials and Methods: Sixteen survey sets (eight languages × clinical vs non clinical) yielded 387 Likert ratings (1–5) for uterine fibroid embolization (UFE), transarterial chemoembolization (TACE), and radiofrequency ablation (RFA). Languages (Mandarin, Cantonese, Spanish, Arabic, French, Tamil, Telugu, Korean) collectively represent ≈2.1 billion native speakers (~26% of the world population). To avoid double counting within Sinitic languages, unique coverage is estimated once; total potential reach (native + L2) is ≈2.6–3.0 billion (~32–37%). Primary endpoints were pooled mean ratings with 95% confidence intervals and Top 2 (good/excellent) proportions. Procedure level differences were examined within languages. Between language differences were tested using Kruskal–Wallis with Holm adjusted pairwise comparisons. Qualitative comments underwent thematic analysis.

Results: Clinical reviewers rated translations as highly appropriate (pooled mean ≈4.6–4.7/5). Non clinical participants demonstrated strong comprehension (pooled mean ≈4.5–4.6/5) with overall Top 2 ≈95%. No material within language differences were observed across UFE, TACE, and RFA. Kruskal–Wallis tests across languages showed no statistically significant disparities for clinical or non clinical outcomes; however, Holm adjusted pairwise contrasts indicated lower non clinical comprehension for French versus Spanish and for Tamil and Cantonese relative to higher scoring languages (notably Telugu/Korean) (p< 0.05). Qualitative findings aligned with these patterns, highlighting strong clarity/cultural fit for Spanish/Arabic and Korean/Telugu, recommended stylistic refinement for French, terminology/register corrections for Tamil, and a preference for proper written Mandarin for print alongside an RFA phrasing correction (emphasizing thermal tumor ablation) for Cantonese.

Conclusion: GPT-4 generated, non-post edited IR instructions achieved high clinical appropriateness and lay comprehension across eight languages spanning roughly one quarter of the world’s native speakers and one third of potential readers.