Abstract
Artificial intelligence (AI) is no longer a distant promise in rheumatology; it is already reshaping how we interrogate data and understand rheumatic diseases. The articles gathered in this special collection, Artificial Intelligence in Rheumatology: Opportunities, Challenges, and Future Directions, offer a panoramic yet pragmatic view of what is happening now and what will soon be possible in the intersection between AI and rheumatology.
A national survey among German rheumatologists opens the collection and reveals a relevant paradox. 1 Respondents admire AI for its potential to accelerate diagnosis and lighten administrative load, yet relatively few have adopted it at the bedside. The barriers they cite, including patchy training, medicolegal uncertainty and anxiety over data security, highlight that the largest obstacles are not technical but human. Encouragingly, the same clinicians express a strong appetite for structured education, suggesting that carefully designed training programmes could turn scepticism into skilful use. The survey makes clear that the perceived bottlenecks are governance and accountability rather than model performance; this argues that AI readiness in rheumatology is now primarily an organizational and regulatory task rather than a purely technical one. A structural reason that hinders implementation is regulatory availability: current FDA authorizations for AI-enabled software cluster in radiology and cardiology, with no tools specifically indicated for rheumatology. There are some examples of CE-marked applications in the European Union relevant to rheumatology, such as the ROPCA platform. It combines a robotic ultrasound scanner (ARTHUR) with an AI model (DIANA) to autonomously capture and score synovitis in hand joints according to EULAR-OMERACT criteria; DIANA V2.0 is CE-certified under the Medical Device Regulation (MDR) as a class IIa device and has been deployed in European rheumatology clinics. 2
Several contributions then demonstrate what is already achievable when robust data meet thoughtful methodology. Analysing the REGISPONSER registry, investigators used cluster analysis to expose how socioeconomic deprivation magnifies disability in radiographic axial spondyloarthritis (r-axSpA), even in an era of widespread biologic access. 3 Their research suggests that the socioeconomic status of patients with r-axSpA may have implications for disease severity and permanent disability, despite the similar use of drugs. Another study, spanning the REGISPONSER and RESPONDIA cohorts, compared 182 machine-learning pipelines before selecting a simple decision tree to predict inflammatory activity in anti-TNF-treated patients, showing that activity levels appear strongly influenced by quality-of-life indicators. 4 Notably, the model selected shows that clinically interpretable approaches built on routine registry data can already prioritize patient-reported quality-of-life indicators as dominant signals of inflammatory burden under anti-TNF therapy. The authors’ deliberate preference for transparency over marginal gains in accuracy is a timely reminder that interpretability remains indispensable when lives, rather than click-through rates, are at stake.
Natural-language processing (NLP) extends this spirit of curiosity to the unstructured realm. By applying BERTopic to two decades of rheumatology abstracts, one group reconstructs the evolution of our field in 45 topics, from enduring fascinations with rheumatoid arthritis (RA) and systemic lupus erythematosus to the recent surge of interest in JAK inhibition and spinal surgery. 5 Such panoramic mapping can guide funding bodies and young investigators alike towards areas ripe for impact. The conceptual sophistication of NLP methodologies is further exemplified by the explanation of Retrieval-Augmented Generation (RAG) as a promising evolution in AI’s integration into rheumatology. 6 Addressing critical limitations of large language models, such as hallucinations and inaccuracies in specialized domains, RAG combines real-time information retrieval with generative AI to significantly enhance output accuracy and context relevance. Because RAG maintains an explicit link between each generated answer and its supporting source, it enables auditability and rapid updating. This can yield more precise, context-aware results, offering a reliable foundation for clinical decision support systems. Beyond static prediction, truly autonomous therapeutic loops, common in diabetes technology, have not yet entered routine rheumatology practice. 7
A thorough review further widens the lens, detailing AI applications across RA, axSpA and psoriatic arthritis (PsA). 8 In RA, machine-learning models already help clinicians anticipate methotrexate response and flare risk, while wearables translate joint stiffness into continuous data streams that patients can collect at home. In axSpA, automated reading of sacroiliac magnetic resonance imaging and computed tomography scans may shorten the diagnostic delay that has impacted this disease for decades. PsA research shows equally rapid progress: AI systems now sift electronic health records to detect subtle cutaneous signals that foreshadow musculoskeletal involvement, enabling earlier referral from dermatology to rheumatology. The review also makes the point that the next decisive step will be multimodal integration, by linking imaging, genetics and real-world data. This evolution will only be acceptable to clinicians and regulators if three conditions are met: high data quality, interpretability of model reasoning and explicit ethical oversight. Without these safeguards, AI risks remaining a patchwork of promising pilots rather than a reproducible standard of care across RA, axSpA and PsA.
Taken together, the collection paints a vivid but balanced picture. AI can illuminate hidden inequities, compress diagnostic timelines and personalize therapy, yet its success hinges on people: the clinicians and researchers who must trust it, the patients who must benefit and the regulators who must protect them. Education, transparency and equity are therefore not afterthoughts but the very engines of progress. As rheumatologists, we should approach AI not as a black-box oracle but as a collaborative colleague; one that will help us deliver care that is faster, fairer and, above all, more human.
