Patient experiences, attitudes, and profiles regarding artificial intelligence in rheumatology: a German national cross-sectional survey study

Artificial intelligence (AI) has the potential to transform all stages of the patient journey [1]. Especially large language models (LLMs), such as ChatGPT, may accelerate the implementation of AI in rheumatology by rapidly transforming the traditional patient–physician relationship into a digital health triad, complemented by AI. Recent studies have demonstrated that LLMs can match rheumatologists in diagnostic accuracy [2] and, in some cases, outperform approved and established medical diagnostic support systems [3]. They are increasingly capable of providing safe treatment recommendations across a broad spectrum of rheumatic diseases [4]. Notably, for patient questions related to general medicine and also systemic lupus erythematosus (SLE), LLM-generated answers were rated as more accurate, higher in quality, and more empathetic than those provided by human experts [5, 6]. These initial findings underscore the transformative potential of LLMs in augmenting rheumatology care.

However, the successful integration of AI into routine care depends not only on its technical performance but also on patient acceptance, trust, and willingness to engage with these technologies. While the technical capabilities of AI in rheumatology have been increasingly studied, these evaluations have typically included rheumatologists [7] and relied on their assessments [8]. To our knowledge, few studies have examined patient perspectives on AI specifically within rheumatology.

This study aimed to address this critical gap by conducting a nationwide web-based survey to explore rheumatic patients’ experiences, attitudes, and expectations toward AI. The results offer new insights to guide the development of AI tools in rheumatology and to shape healthcare policies and practices that promote patient-centered and responsible AI use.

An initial draft of the questionnaire in German was collaboratively developed by four rheumatologists (JK, HL, MK, AH) based on a comprehensive literature review. To ensure content and face validity, the draft was additionally reviewed by two patient research partners (GB, CEA) affiliated with the German League Against Rheumatism. Based on their feedback, both the content and wording of the survey were revised.

The final questionnaire included 18 questions (English translation, see Supplementary File 1). Participation took approximately 15 min per participant. The questionnaire collected demographic data (5 questions) including age, sex, education level, and type of rheumatology treatment center. Participants were also asked to indicate their specific rheumatic disease. In addition, the survey assessed prior experiences with AI in a medical context (3 questions), as well as attitudes regarding its use in healthcare, including interest, willingness to use, and potential concerns (10 questions).

An anonymized cross-sectional web-based study was conducted via REDCap (Research Electronic Data Capture; Vanderbilt University), hosted by the University of Marburg, Germany, and was accessible between March 25 and May 30, 2025. Adults currently receiving care from a rheumatologist were eligible to participate. The survey was distributed through QR codes in German university and non-university rheumatology clinics and promoted by the German League Against Rheumatism via its newsletter and website. The Ethics Committee of Philipps-University Marburg confirmed that formal ethical approval was not required for this anonymous survey study (reference: 25–89 ANZ). No formal sample size calculation was performed, as this study followed an exploratory design.

Results were reported according to recommendations by Zimba and Gasparyan [9]. Parametric statistical methods were applied based on the central limit theorem. Numerical variables were summarized as means and standard deviations, while categorical variables were described using absolute and relative frequencies (percentages). Group differences in survey responses by clinical and demographic characteristics were analyzed using t-tests with Cohen’s d and one-way analysis of variance (ANOVA) with Eta squared following Tukey’s HSD post hoc tests.

To identify distinct patient profiles based on patients’ responses to AI-related items, an unsupervised cluster analysis was conducted. Only complete data sets (N = 739) were included. Prior to clustering, relevant variables were standardized. The data for the variables AI usage, AI-supported second opinion, providing data for AI research, self-management, and disease monitoring were recoded so that higher values indicated stronger usage or more positive assessments. Scores for individual dimensions with multiple sub-dimensions were calculated by taking the row-wise mean of the relevant sub-dimension variables (potential AI applications, perceived advantages, facilitators and barriers). Dimensionality reduction was performed using Principal Component Analysis (PCA) to explore the data structure and visualize group separation. K-means clustering was applied, and the optimal number of clusters was determined using the elbow method [10]. Differences between clusters were assessed using one-way ANOVA followed by Tukey’s HSD post hoc tests. Associations between cluster membership and demographic or clinical variables were examined using Fisher’s exact tests (pairwise, where appropriate). In addition to p-values, effect sizes were calculated to quantify the magnitude of observed differences. Eta squared (η²) was reported for continuous outcomes such as AI interest, perceived usefulness, and AI usage. Effect sizes (η²) were interpreted as small (≈ 0.01), medium (≈ 0.06), and large (≈ 0.14). Cramér’s V was used to measure the strength of association between categorical variables (e.g., sex, age group, education, healthcare setting, and disease) and cluster membership. All analyses and figures were generated in R (version 4.5.0; RStudio 2025.05.0) using the packages FactoMineR, ggplot2, and cluster. Figures were generated in R using ggplot2 and factoextra.

For the Sankey diagram, categorical data were preprocessed in R and then exported to the online tool SankeyMATIC (https://sankeymatic.com/) for visualization.

This study presents novel findings from the first national patient survey that systematically explores the use, perceptions, and expectations of AI among individuals with rheumatic diseases. These findings complement recent reviews [8] of AI use cases and developments in rheumatology and add the patient perspective to clinician-focused studies, such as the 2024 national survey of German rheumatologists conducted by Holzer et al. [7].

Notably, 27% of both patients and rheumatologists [7] reported using AI for health-related purposes. Consistent with earlier studies identifying information access as a top priority among patients [11, 12], AI in our survey was most commonly used to obtain health-related information. This aligns with data showing that only 24% of German rheumatology patients feel well-informed about their disease [13], highlighting the potential of AI to support patient education. Tools like www.lupusgpt.org, which use retrieval-augmented generation, are increasingly adopted and may offer information that is not only accessible but also comparable in quality to rheumatologists [5, 14].

Patients also showed strong interest in AI-supported diagnostics and endorsed its use by physicians for second opinions. Patients increasingly consult symptom checkers or online sources before medical appointments [12, 15]. These behaviors, along with persistent diagnostic delays [16], suggest opportunities for AI to improve early detection and referral. Recent studies show that off-the-shelf LLMs outperform existing decision support systems in diagnostic accuracy and speed [3]. A randomized controlled trial further demonstrated that medical students performed significantly better solving rheumatology vignettes when supported by ChatGPT [17]. While further research is needed, rheumatology may soon follow other specialties, such as oncology, where AI-supported screening is already routine.

In our study, the majority of rheumatology patients supported the use of AI-generated second opinions by doctors consistent with the results of a study involving breast cancer patients [18]. However, the findings from Reis et al. [19], which indicated that the public perceives fictional doctors using AI as less competent, less trustworthy, and less empathetic, seem to contradict our results.

Our exploratory clustering approach confirmed that the vast majority of patients show a certain affinity for AI and are at least not opposed to its future use in rheumatology, which is consistent with data on the German general population [20] and with results of a multinational survey on attitudes toward AI among hospital patients [21]. Only 102/739 (13.8%) patients clustered in the AI-skeptical group. More experience with AI in a medical context is likely to change patients’ perspectives in the future. The fact that 37.9% of study participants stated they “did not know about AI” is an important finding of our study, clearly demonstrating the need for education.

The identification of three distinct patient clusters confirms the heterogeneity of AI attitudes. In line with previous findings on digital health literacy [11], younger and more educated patients were more likely to belong to the AI-savvy group. This finding that age and education are linked to AI attitudes is important, as it suggests that implementation strategies should account for these demographic differences. Specifically, tailored patient education and communication strategies may be needed to engage older or less educated patient groups and ensure equitable understanding and acceptance of AI-based health technologies.

Patients with fibromyalgia reported lower current use of AI compared to those with other rheumatic diseases. The underlying reasons for this observation warrant further investigation in larger, dedicated studies, given the small sample size and modest effect. When implementing AI in rheumatology, it will be crucial to pay special attention to these groups to prevent further exacerbation of health inequalities.

Although patients perceived multiple benefits, they also voiced concerns about data privacy, AI reliability, and reduced human interaction. The fact that nearly three-quarters identified faulty AI as a major barrier underlines the importance of ensuring transparency, reliability, and clinical validation of AI systems before their widespread implementation. Concerns about reduced human interaction suggest that patients value the interpersonal aspects of care and may resist AI tools perceived as replacing, rather than supporting, clinicians. Furthermore, the notable proportion citing data privacy and trust issues emphasizes the need for clear communication about data protection, algorithmic accountability, and the complementary role of AI in shared decision-making. These concerns also proved to be significant in other studies involving non-rheumatological study participants [22‐24]. Addressing these issues through transparent information, patient involvement in AI development, and clinician–patient dialogue will be crucial for fostering trust and acceptance. Scientific validation, physician endorsement, secure data protection and user-friendliness were key factors influencing acceptance, underscoring the important role of clinicians and academic institutions in guiding responsible AI integration. The finding that over one-third of patients regarded recommendations from patient organizations as a facilitator shows the critical role of such organizations when implementing AI in rheumatology.

A major strength of this study is its multicenter design, including university and non-university rheumatology centers and a broad range of rheumatic diseases. The direct involvement of patient representatives throughout all study phases further strengthens the relevance of the findings. Nonetheless, the online format and QR code access likely introduced a considerable selection bias, favoring digitally engaged and higher-educated individuals which may have biased the results toward greater AI acceptance. Additionally, as the survey was conducted in Germany, the findings reflect the country’s specific digitalization culture, which may limit their generalizability to other healthcare systems. Patients with a university degree were slightly overrepresented compared to data on the educational level of the general German population [25], which may have influenced representativeness of the results. A further methodological limitation is the use of self-reported diagnoses.

PCA and cluster analysis were employed to identify and visualize patient clusters representing distinct attitudes and experiences related to AI use. Both techniques are widely applied for dimensionality reduction and pattern identification, yet they entail certain methodological limitations. PCA assumes linear relationships among variables and may result in information loss during dimensionality reduction. As with most clustering methods, the cluster analysis results depend on the type and number of variables included, meaning that changes in variable selection can produce different cluster solutions, even within the same dataset. Despite these constraints, PCA provides an efficient means of summarizing complex datasets while retaining most of their variance, and cluster analysis reveals latent group structures that enhance and clarify data interpretation.

This study highlights a clear gap between patients’ current use of AI for health-related purposes and their interest in AI-supported medical applications. Patients and rheumatologists share similar priorities, particularly around diagnostics and education. Realizing the full potential of AI in rheumatology will require rigorous validation, targeted implementation strategies, and a proactive role from the rheumatology community in ensuring that AI serves both clinical and patient needs effectively.