RETURN TO ABSTRACT LISTING


Objective:
Nearly 1 in 5 people living in the United States speak a language other than English at home. With increasing globalization and diversity, it is imperative to ensure that access to healthcare grows as well. It has been previously reported that patients with limited English-speaking proficiency face a variety of healthcare disparities. Recent reports have shown ChatGPT’s effectiveness at translating radiology reports into simplified English. This study was designed to explore the capabilities of ChatGPT for the purpose of simplifying and translating radiology reports into Spanish, Hindi, and Russian languages, exploring its ability to improve accessibility in healthcare.

Materials and Methods:
In this institutional review board approved study, 50 random abdominal computed tomography (CT) reports were collected. Deidentified reports were fed to ChatGPT (4.0) with a tailored prompt instructing it to simplify (and translate) the report to a different language – Spanish, Hindi, Russian and English. The translations were done iteratively, in separate instances. The processed reports were evaluated by board eligible radiologists proficient in the specific language and rated on factual correctness, potential harmful errors, completeness, and explanation of medical terms, each graded on a 5-point scale. The Spanish, Hindi and Russian reports were also rated on the quality of translation. The mean scores in each category were compared between the Russian, Spanish and Hindi languages using Kruskal-Wallis test. Each translated version was also compared with English version in the first four categories using Mann-Whitney test. Potential uses, scopes for improvement and pitfalls were discussed.

Results:
The Spanish translation outperformed the Hindi and Russian version significantly in categories 1 (factual correctness) and 3 (completeness). All the translated versions performed significantly worse compared to the English version in category 4 (explanation of medical terms). Notably, the Hindi translated version performed significantly worse in all 4 categories. The Russian translated version was also significantly worse in the third category. In the first three categories, the Spanish translation, and the Russian translation in the first two categories demonstrated no statically significant difference from the English version. The lowest scores were related to grammar and difficulties translating medical context and English verbiage. Typographical errors in the original reports negatively affected the translation.

Conclusion:
Improving multilingual accessibility in health is a promising new use case for natural language processing applications. Our study demonstrates ChatGPT’s potential ability in this field, translating reports and communicating pertinent clinical information with limited errors. Although the translations were not perfect, and more training and tailoring will be required for languages that are not as commonly used in medical literature, this demonstrates the potential this avenue holds with further advances.