S2670
RTT - Patient experience and quality of life
ESTRO 2026
cancer is accurate for patient self-education. Material/Methods:
A multi-pronged approach was used to generate a set of common prostate cancer-related prompts. These prompts consisted of frequently asked questions gathered from patient information websites, as well as prompts provided by a Clinical Oncologist specialised in the treatment of prostate cancer. These were inputted into two different LLMs: ChatGPT-4 and Gemini 1.5. Prompts were inputted both consecutively (one after another) and in isolation (in a new chat each time) to investigate the impact of these alternative approaches on the information provided. Responses generated by the LLMs were evaluated for agreement with the European Association of Urology (EAU) guidelines using a 5-point concordance scale, which acted as a surrogate for the level of accuracy of LLM responses. Data analysis included descriptive statistics and Mann-Whitney U tests to identify discrepancies in information accuracy, as well as highlighting areas relating to prostate cancer which were answered poorly. Results: The accuracy of both LLM responses assessed was generally good with 92.6% of total responses showing at least ‘some agreement’ with the EAU guidelines; scoring ≥ 3 on the concordance scale. Forty three out of 68 responses (63.2%) received a score of ≥ 4 indicating agreement or strong agreement with the guidelines (Figure 1). A few inaccuracies were identified (5 out of 68 responses scoring ≤ 2), primarily related to treatment options. Complex or double- barrelled prompts were notably lower in accuracy compared to those inputted in simpler formats. No significant difference (p>0.05) in accuracy was found between responses inputted consecutively and in isolation (Figure 2).
Conclusion: The ongoing possibility of generating inaccurate information limits the use of LLMs as an independent source for patient information. The limitations associated with LLMs, such as the importance of user expertise and decreased accuracy in complex scenarios, represent a disadvantage surrounding patient use. Despite this, the potential of LLMs continues to increase as the technology improves, and further development of oncology specific models could offer an improved solution for patient education. Keywords: artificial intelligence, patient education Neck fibrosis in Head and Neck Cancer is under- recognised and lacks rehabilitation: Combined findings of a cohort study and a national survey. Rachel Wijayarathna 1,2 , Patrick Corbett 3 , Sabina Khan 4 , Boris Tocco 5,2 , Imran Petkar 2 , Delali Adjogatse 2 , Teresa Guerrero-Urbano 2 , Anthony Kong 2 , Mary Lei 2 , Gareth D Jones 1,5 , Nicola Peat 1 , Ruheena Mendes 4 , Miguel Reis- Ferreira 2,6 1 Physiotherapy, Guys & St Thomas' NHS Foundation Trust, London, United Kingdom. 2 Department of Oncology, Guys & St Thomas' NHS Foundation Trust, London, United Kingdom. 3 Ageing & Health, Guys & St Thomas' NHS Foundation Trust, London, United Kingdom. 4 Department of Oncology, University College Hospital, London, United Kingdom. 5 Centre for Human & Applied Physiological Sciences (CHAPS), Faculty of Life Sciences & Medicine, King's College London, London, United Kingdom. 6 Centre for Host- Proffered Paper 2825
Made with FlippingBook - Share PDF online