Background: Systemic lupus erythematosus is a rare multisystemic disease that often requires lifelong treatment and affects various aspects of life. Patient education is an essential pillar of disease management but is often impaired by physician’s lack of time and staff. To provide high quality patient information, the initiative Lupus100.org was launched[1]. Here, lupus experts and Lupus Europe listed and answered the 100 most important questions on lupus. The creation of such information collections needs a lot of time and resources. With the wide availability of large language models (LLM) such as ChatGPT, the question arises to what extent they could support physicians in the care of patients.

Objectives: To assess the capability of the LLM ChatGPT-4 vs. physician-generated responses to answer the 100 most frequently asked patient questions related to lupus.

Methods: ChatGPT-4 responses were generated by entering the English questions from https://lupus100.org/ in a fresh session on October 16 ^th , 2023. Three senior rheumatologists who were blinded concerning authorship evaluated responses from ChatGPT-4 and Lupus100 independently. The evaluation criteria were quality, empathy (Likert scale 1-5 each) and the selection of a preferred answer. Differences between the scores were analysed using a two-tailed Student’s t-test. The relationship between quality and empathy scores and word count was conducted using Spearman’s Rank Correlation Coefficient. A one-sample Chi-Square test was performed to assess whether there was a preferred source for the answers. Additionally, rates of responses below important thresholds (quality “poor” or “very poor”, empathy “not empathetic”) were compared. All statistical analyses were conducted in SPSS, the significance threshold used was p <0.05. Ethical approval was granted by the ethical committee of the Philipps University Marburg, Germany (23-300 ANZ).

Results: Across the 100 questions, evaluators scored the mean quality of ChatGPT-4 significantly higher than that of Lupus100.org (t = 4.25; p = 0.001) (Table 1). Empathy ratings were comparable between both sources (t = 1.14; p = 0.26). Very few answers were rated as of poor quality or empathy (Figure 1). The preferred response was significantly more likely to come from ChatGPT-4 (57% vs. 43%; χ ² = 5.88, p = 0.02).

Mean (SD) word count of replies was significantly lower for Lupus100.org than for ChatGPT-4 (241 [135] vs. 372 [52]; t = 9.08; p = 0.001). Word count was positively correlated with quality and empathy in Lupus100.org (r = 0.52; p = 0.001 and r = 0.34; p = 0.001). If only answers longer than median word count by Lupus100.org were analyzed, evaluators rated the two sources as equivalent in terms of quality (t = 0.65; p = 0.51) and Lupus100.org significantly better in terms of empathy (t = -2.17; p = 0.03).

Conclusion: In this study, ChatGPT-4 was able to generate responses with high quality and empathy to patient questions concerning lupus. The length of the response seems to influence these parameters and can be produced faster by LLM, while physicians can probably optimize empathy by a thorough edit of answers. Collaboration between LLM and physicians has the potential to improve the availability of high quality and empathic patient information in times of physician shortage.

REFERENCES: [1] Lupus100.org [Internet]. [cited 2023 Dec 27]. Available from: https://lupus100.org/en

Acknowledgements: NIL.

Disclosure of Interests: Isabell Haase Abbvie, AstraZeneca, Boehringer Ingelheim, Galapagos, GSK, Janssen, Lilly, Medac, Novartis, UCB, AbbVie, Boehringer Ingelheim, Medan, AbbVie, Celgene, Chugai, Hexal, Janssen, Medac, UCB, Tingting Xiong: None declared, Antonia Rissmann: None declared, Johannes Knitza: None declared, Julia Greenfield: None declared, Martin Krusche Novartis, Abbvie, Lilly, Galapagos, Pfizer, Medac, GSK, Novartis, Roche, Abbvie, Lilly, Galapagos, Pfizer, Medac, Novartis, Sobi, Sanofi.