ESTRO 2025 - Abstract Book

S2418

Interdisciplinary – Other

ESTRO 2025

To assess the effectiveness of a fine-tuned OpenAI GPT-4o model in providing accurate and contextually relevant answers tailored to Polish oncology settings. Additionally, to outline the protocol for a prospective observational study evaluating the application of the model in clinical care. Material/Methods A database of frequently asked questions and evidence-based answers was developed using Polish oncology guidelines and clinical observations. GPT-4 was fine-tuned using this dataset to enhance its ability to address oncology-specific queries. Fifty sample questions were posed to both the fine-tuned and standard GPT-4 models. Two independent investigators evaluated the responses, assessing quality, relevance, and adaptation to Polish oncology settings. The results informed the design of a prospective observational study involving at least 50 patients undergoing radiotherapy. Results In the preliminary evaluation, the fine-tuned GPT-4o model provided superior answers in 44 out of 50 cases (88%) compared to the standard model. These findings highlight the enhanced capability of the fine-tuned model to address patient queries in a manner consistent with local clinical practice. Based on these results, a prospective study was designed to validate the model’s real-world utility. This study will involve patients using a mobile application incorporating the fine-tuned GPT-4 model. Patients will review the model's responses for accuracy, and patients will evaluate the application’s utility and empathy using a Likert scale post-treatment. Conclusion Preliminary results demonstrate that fine-tuning of GPT-4o for localized clinical contexts significantly improves its performance in oncology-specific applications. The planned prospective observational study will validate the model’s effectiveness in supporting patient education and communication during radiotherapy. These findings provide a foundation for broader implementation and potential commercialization of AI-driven tools in radiation oncology care. The study is funded by the Ministry of Science and Higher Education in Poland, grant number SKN/SP/601573/2024.

Keywords: AI, patient assistance, radiation oncology