Using AI in Forward-Backward Translation of Questionnaires for Men Invited to Prostate Cancer Screening: Methodological Study

Background: Translation is important in research to ensure cultural relevance, accuracy, and generalizability, particularly in cross-cultural studies. The forward-backward translation method of the World Health Organization (WHO) is commonly used to improve linguistic and conceptual accuracy but is often time-consuming and resource intensive. The development of advanced artificial intelligence (AI) offers new opportunities to make the translation process more efficient, potentially reducing time and costs. However, concerns remain regarding the ability of AI to capture cultural nuances and complex linguistic structures, which may affect translation quality. Therefore, evidence on how AI can be effectively integrated into established translation frameworks remains limited. Objective: This study aimed to explore the use of AI in the forward-backward translation process for questionnaires. Methods: We used an adapted version of the WHO 4-step forward-backward translation method to translate the questionnaires from English into Polish. The questionnaires included the Prostate Cancer Screening Education (PROCASE) Knowledge Index, the Attitude Scale, Risk Perception items, and the Brief Health Literacy Scale for Adults. First, 2 AI tools (ChatGPT [GPT-3.5] and Microsoft Bing Copilot) were used for translating from English to Polish. Second, 2 native Polish speakers focused on content understanding independently reviewed and corrected the AI-generated Polish version and agreed on a new version. Third, the AI-generated Polish translation was back-translated from Polish into English using the same AI tools. Any discrepancies were discussed by an expert panel consisting of native speakers of English and Polish. This procedure ensured linguistic accuracy and conceptual similarity. Finally, 3 individual cognitive interviews were conducted with native Polish-speaking men to identify whether the questionnaires measured the intended constructs and to find any issues that the respondents might encounter during the response process. Results: Minor discrepancies between the two AI-generated Polish phrases “umiera z innej przyczyny” and “umiera z powodu innych przyczyn” were merged by native Polish speakers in the PROCASE Knowledge Index. The original questionnaires and the AI-generated questionnaires had minor differences, but they did not affect the meaning of the questions or what was being asked. We conducted individual cognitive interviews (n=3) with participants aged 47 to 74 years. After the interviews, the questionnaires were adjusted with a few changes to make them easier to understand. In the Attitude Scale, the AI-generated Polish translation was changed from “nieco” to “trochę” to align with everyday language and improve understanding. Conclusions: AI can be an effective tool in the translation process, offering time and resource savings while maintaining accuracy. However, human involvement is still needed to optimize translation.