Introduction

Research suggests nearly one in ten individuals is affected by an autoimmune or rheumatic disease – conditions that largely lack definitive cures and show a rapidly rising incidence1,2. In light of these pressing challenges, recent npj Digital Medicine articles highlight how recent developments in artificial intelligence, particularly in generative artificial intelligence (GenAI), represent a potential paradigm shift in autoimmune and rheumatic disease care3,4,5. Traditional AI applications for rheumatic diseases have mainly focused on narrow tasks, like classification or prediction6. In contrast, generative models, such as large language models (LLMs), can produce contextualized clinical content and recommendations informed by large medical datasets, supporting a wide range of clinical tasks6,7. These models may be particularly valuable for rheumatologic practice, which is often defined by diagnostic uncertainty, varied disease presentations, and the need for individualized treatment8,9. Despite this alignment, the applications of this technology across autoimmune and rheumatologic clinical care, research, and administrative workflows remain largely unexplored.

AI-enhanced clinical decision-making

Generative AI demonstrates early potential to improve diagnostic accuracy, treatment guidance, and clinical decision-making in rheumatic disease management. GenAI’s potential to advance clinical care partly lies in its ability to combine multiple types of clinical information; by integrating patient symptoms, laboratory data, imaging findings, and even genomic profiles, these systems can generate clinical insights that inform decision-making in real-time3,7. These systems show particular proficiency in navigating the complexity inherent to rheumatic diseases, where diagnostic and therapeutic decisions often require the synthesis of clinical elements across multiple organ systems and timeframes9,10. A recent validation study demonstrated that foundational LLMs achieved high diagnostic accuracy in inflammatory rheumatic diseases, identifying a higher proportion of expert-curated cases correctly compared to human specialists11. Similarly, GPT-4’s performance in musculoskeletal radiology interpretation matched the diagnostic accuracy of radiology residents when provided with both medical history and imaging findings, highlighting its multimodal capabilities12. These platforms also excel in rare disease identification, with select LLMs achieving high accuracy in diagnosing uncommon and orphan diseases – a valuable capability given the rarity of many rheumatic diseases that may challenge even experienced clinicians13,14. GenAI and related advancements have also demonstrated the ability to incorporate dermatologic findings to enhance the evaluation and management of autoimmune diseases, as in diagnosing lupus or evaluating alopecia areata severity15,16. While these early findings are promising, the evidence base remains limited, predominantly derived from curated clinical vignettes and structured datasets rather than the real-world prospective trials and longitudinal studies typically needed for widespread implementation10,11,12.

Platforms have also demonstrated accuracy in providing treatment guidance, with LLMs achieving high accuracy and concordance with guidelines when delivering methotrexate-related information for rheumatoid arthritis patients17. This early evidence suggests their potential to serve as reliable clinical decision support tools that can adapt to differing patient presentations while incorporating evolving evidence-based guidelines. Importantly, real-world evidence shows that clinician-AI collaboration can improve care quality: in a multicenter randomized controlled trial of physicians, LLM assistance increased correct management-reasoning scores without raising harmful decision rates, underscoring how interactive generative AI can translate into improvements in clinical outcomes18. For instance, a GenAI system could potentially evaluate a patient presenting with joint pain and a facial rash, recommend targeted autoantibody testing and evaluation for internal organ involvement including muscle inflammation and kidney function studies. It could then integrate these results to guide further testing to differentiate dermatomyositis from lupus diagnosis and recommend initial treatment, tailored to the individual’s needs (Fig. 1).

Fig. 1
figure 1

Selected applications of generative artificial intelligence in autoimmune disease care.

Full size image

Moving beyond the administrative maze

Multiple studies have suggested that physicians spend nearly twice as long on administrative work as on face-to-face care19,20. The administrative burden inherent to rheumatic disease management, characterized by extensive documentation and care coordination requirements, can outweigh that of other specialties, and presents a compelling target for generative AI optimization21. Current LLMs demonstrate significant capability in reducing documentation overhead through automated generation of clinical notes, discharge summaries, and administrative communications, with real-time suggestions that maintain clinical accuracy while improving efficiency22,23,24,25.

These systems also show particular value in patient engagement through the generation of personalized educational materials that can be tailored to individual health literacy levels, cultural preferences, and disease-specific considerations – a critical capability given the complexity of rheumatic diseases and their treatments26,27,28.

Additionally, generative models have the potential to streamline insurance authorization processes by automatically creating prior approval documentation and translating complex clinical rationales into formats required by payers, potentially reducing delays in accessing specialized rheumatologic treatments, like biologics29,30. Future implementations are anticipated to integrate ambient listening technologies that can automatically capture and structure clinical encounters, convert conversational exchanges into structured documentation, and provide real-time clinical decision support, thereby allowing clinicians to focus more directly on patient care rather than administrative tasks31,32.

Accelerating discovery

Generative AI may help in reshaping the research landscape in rheumatic diseases by enabling novel approaches to drug discovery, clinical trial design, and scientific hypothesis generation.

In pharmaceutical development, transformer-based models have been successfully adapted for molecular representation learning, enabling the identification of novel therapeutic targets and prediction of molecular properties essential for drug development in rheumatic diseases33. Additionally, emerging generative drug repurposing technologies may potentially enable identification of new therapeutic applications for existing medications in autoimmune diseases34. These technologies can also help create digital twins - sophisticated patient simulation models trained on clinical trajectories, laboratory results, and treatment responses35. For instance, these simulations can reproduce the complex cellular interactions driving rheumatoid arthritis, enabling testing of drug mechanisms, prediction of disease trajectories, and identification of new therapeutic targets36. Similarly, GenAI can learn from existing patient data to generate synthetic patient profiles; by expanding small cohorts of rare diseases such as lupus, systemic sclerosis, and vasculitis, these profiles could provide the statistical power needed to compare treatments and predict patient outcomes with more confidence37,38.

These systems also demonstrate the capacity to aid in medical research beyond data analysis: generating research hypotheses, conducting literature reviews, creating cloud-based data pipelines that allow secure and remote analysis of sensitive patient data, and assisting in manuscript preparation, thereby potentially accelerating the pace of scientific discovery while maintaining methodological rigor39,40,41,42.

Challenges for AI in rheumatic disease care

Significant challenges remain across multiple domains for the successful implementation of generative AI in advancing clinical care, that are briefly alluded to here.

Clinical validation is a particular challenge given the rarity of many rheumatic diseases; limited patient populations often make it difficult to create robust trials and generate sufficiently large validation datasets for AI systems43. The heterogeneity in study designs and outcome measures has been further complicated by inconsistent reporting standards, prompting the recent development of LLM-specific reporting guidelines to improve research transparency and reproducibility44. To address performance challenges, generative AI approaches can include fine-tuning methods for domain adaptation, synthetic data generation to expand limited datasets, or alternative architectures such as diffusion or adversarial networks27,42. Of particular relevance for clinical applications is retrieval-augmented generation (RAG), which integrates external knowledge sources (e.g., clinical guidelines or verified resources) into model outputs to improve factual accuracy and mitigate hallucinations - representing a potential pathway toward safer and more reliable use in medicine45.

Model interpretability and explainability remain critical barriers, as rheumatologists may require transparent reasoning processes to maintain clinical confidence and ensure appropriate care, particularly given the complexity of multi-organ diseases where treatment decisions must account for numerous interacting factors including disease activity, medication interactions, and patient preferences46,47. While studies have demonstrated that LLM-generated explanations may sometimes represent post-hoc explanations rather than an actual account of the model’s internal reasoning, they may still offer value: flaws in these explanations, such as contradictions, omissions, or unsupported logic, can correlate with inaccurate or biased outputs and provide a flag for clinician oversight46. However, evidence remains preliminary and mixed, with some clinician–machine learning collaboration experiments noting that explanations accompanying incorrect recommendations can heighten over-reliance and reduce clinician accuracy48,49,50.

Data integration challenges may be particularly pronounced in rheumatology, where patient care often involves multiple subspecialists, fragmented electronic health records, and longitudinal disease monitoring that spans decades46,51. Ensuring secure, interoperable data sharing requires robust technical infrastructure and standardized data formats that many healthcare systems currently lack – particularly relevant given the heterogeneous laboratory and serologic testing prevalent for autoimmune diseases52. Regulatory oversight of non-deterministic AI systems remains nascent, with current frameworks emerging to address the unique challenges of generative models that produce variable outputs - particularly concerning when applied to high-stakes clinical decisions regarding immunosuppressive therapies53.

Finally, successful implementation ideally requires fluid integration into existing clinical workflows and training programs to ensure rheumatic disease providers can effectively interpret and utilize AI-generated outputs45,47.

Conclusion

Generative AI does not substitute for the human expertise that arises from years of prior experience and personal interaction with patients. However, collectively, these advances position generative AI to potentially transform rheumatologic practice toward greater precision, efficiency, and patient-centered care - provided its integration proceeds with rigorous validation, transparency, and equitable oversight.