人工知能による英語およびスペイン語の手術同意書の簡略化：1サイズはすべてに適合しない

Guidelines recommend patient-facing material be written at or below a 6-9 th-grade reading level; however, surgical consent forms frequently exceed this standard. Artificial intelligence chatbots offer accessible tools to enhance readability, but their performance for non-English material remains unclear.

We evaluated three strategies to improve consent readability. First, ChatGPT-4.0 was prompted to simplify pre-existing surgical consent forms in English and Spanish. Second, a custom generative pre-trained transformer (GPT) was trained using these outputs and tested on a separate validation cohort. Finally, both GPT's were prompted to translate English forms into Spanish, targeting a 6th-grade reading level. Readability was measured by the Fry Readability Formula, Simple Measure of Gobbledygook (SMOG) Index, Flesch-Kincaid Grade Level for English, and the Gilliam-Peña-Mountain (GPM) and SOL Readability Formulas for Spanish.

ChatGPT-4.0 modestly improved English readability, with median Fry decreasing from 12.0 to 10.0, SMOG 12.3 to 11.7, and Flesch-Kincaid 9.2 to 8.6 (all P<0.05). Spanish readability did not significantly improve (median GPM 9.5 to 10.0; SOL 9.6 to 9.1; both P>0.05).Our custom GPT improved English readability, reducing median Fry from 12.0 to 7.0, SMOG 12.3 to 9.3, and Flesch-Kincaid 9.6 to 5.9 (all P<0.001). Spanish readability gains were modest with median SOL decreasing from 9.5 to 7.3 (P<0.01). GPM was unchanged (8.0, P>0.99).English-to-Spanish translation outperformed direct Spanish simplification. ChatGPT-4.0 achieved median GPM of 8.0 (P=0.25) and SOL of 7.6 (P<0.001, compared to original Spanish consents). Translation via the custom GPT achieved the largest improvement (median GPM 7.0 and SOL 6.8; both P<0.05, compared to original Spanish consents).

Tailored artificial intelligence chatbots can improve readability of patient-facing material; however, diverse strategies are necessary to adapt existing models for multilingual populations.