Artificial intelligence is reshaping how rheumatologists detect, classify, and manage disease.
Artificial intelligence is moving rapidly from the research lab to the rheumatology clinic.
At ACR Convergence 2025, researchers unveiled a wave of AI-driven tools capable of transforming how clinicians diagnose, classify, and manage rheumatic diseases – from imaging interpretation to disease differentiation and outcome prediction.
In imaging and disease classification, several teams demonstrated how machine learning models could enhance diagnostic precision and reduce variability between readers.
One group introduced an AI system that detects active sacroiliitis lesions on MRI in patients with axial spondyloarthritis, offering a scalable method to standardise interpretation, identify active inflammation with high accuracy and potentially accelerate early diagnosis.
The poster presentation detailed a multinational retrospective study that analysed 684 sacroiliac joint MRIs collected from six international centres. All scans were quality-checked and independently reviewed by experienced radiologists according to the Assessment of Spondyloarthritis International Society (ASAS) criteria, which define active sacroiliitis as bone marrow oedema (BME) visible on two consecutive MRI slices.
Using this rigorously labelled dataset, the researchers trained a deep-learning model to automatically segment the sacroiliac joints and detect BME across slices, classifying scans as ASAS-positive or negative. The system also generated heat maps to visualise the regions driving its predictions, aiding interpretability and clinical confidence.
Overall, 32% of patients met ASAS criteria for active sacroiliitis. The AI model achieved a balanced accuracy of 81.3%, with 71.4% sensitivity and 91.1% specificity. Its performance was even stronger – 89.3% balanced accuracy, 84% sensitivity, and 94.5% specificity – in cases with high-quality images and consensus readings.
However, results declined with poor image quality or subtle lesions, reinforcing the importance of expert oversight in ambiguous cases.
Another team presented ASembleNet, a hybrid transformer–convolutional neural network model that accurately classified ankylosing spondylitis based on MRI scans, demonstrating the growing utility of computer-assisted diagnostics in inflammatory spinal disease.
Developed using thousands of real-world and publicly available MRI images, ASembleNet integrates convolutional and transformer-based architectures to detect subtle inflammatory changes that often elude early diagnosis.
In testing, the model achieved 99% accuracy on coronal MRI and 98% on axial MRI, outperforming established AI models such as Xception, ResNet50, and InceptionV3.
The research team trained the model on more than 4000 MRI images from AS patients and healthy controls collected at Elazig Fethi Sekin City Hospital between 2018 and 2023, supplemented by expert-reviewed images from the Assessment of Spondyloarthritis International Society (ASAS) Online Case Library and normal scans from Radiopaedia.org.
Rigorous preprocessing ensured consistency across data sources, while explainability tools such as Grad-CAM helped confirm that the model’s visual attention corresponded with clinically meaningful regions of inflammation.
External validation with ASAS cases confirmed the model’s strong generalisability, with ASembleNet maintaining high sensitivity and specificity across imaging planes.
The team has also launched a web-based platform allowing clinicians to upload MRI scans and receive real-time classification results, complete with heatmaps indicating areas driving the model’s decision.
“Artificial intelligence is advancing from concept to clinical utility in rheumatology,” said Dr Miral Hamed Gharib, from the Hamad Medical Corporation, Qatar, and the lead investigator for the abstract.
“These findings underscore the power of technology to not only refine diagnoses using imaging and synthetic data but also to uncover critical predictors of disease outcomes that inform treatment strategies.”
Generative AI also featured prominently across the program. Using a CycleGAN and EfficientNetB7 framework, investigators improved hand osteoarthritis classification by generating high-quality synthetic imaging data. Their findings were reported in a poster at ACR.
They said applying the CycleGAN model to generate realistic synthetic images of severe disease, addressed a common problem in AI training – the scarcity of high-quality images representing advanced OA.
Severe cases (KL3 and KL4) make up only about 2% of interphalangeal joints in the Osteoarthritis Initiative dataset, limiting the accuracy of automated classifiers. By transforming abundant KL0 and KL1 images into synthetic KL3 and KL4 cases, the team expanded the dataset and retrained an EfficientNetB7 model.
Incorporating 30% synthetic data into the training set improved KL3 classification accuracy by 6% and KL4 by 3.1%, with minimal effect on milder grades.
The generated images maintained realistic morphology and texture, ensuring that the model learned meaningful disease features rather than artifacts.
“Our study demonstrates the potential of generative AI in improving hand OA classification, particularly for the severe radiographic OA (KL3 and KL4),” the researchers concluded.
“The proposed approach effectively implemented morphological changes to generate OA cases from healthy cases while keeping textural consistency, which leads to improved classification performance.”
In another study presented as a poster, researchers used synthetic data to train an AI model that distinguished seronegative rheumatoid arthritis from psoriatic arthritis in patients without skin psoriasis.
Differentiating seronegative (RA) from psoriatic arthritis (PsA) without skin psoriasis has long challenged rheumatologists, who have often had to rely on subtle clinical features such as axial involvement or enthesitis, which are present in only a minority of patients.
Researchers used data from 200 real patients (105 with PsA sine psoriasis, 95 with seronegative RA) to develop synthetic patient profiles using a Conditional Tabular Wasserstein Generative Adversarial Network (GAN).
The synthetic cohort achieved a 92% statistical fidelity to the real population while maintaining a 73% privacy preservation score, demonstrating that AI can generate realistic yet non-identifiable patient data.
The study also tested the utility of synthetic data in improving machine learning classification. A Random Forest model trained on real data achieved 85.5% accuracy in distinguishing PsA sine psoriasis from seronegative RA.
After incorporating the synthetic cohort into the training dataset, model accuracy increased to 86.2%.
“The first-ever application of generative AI in Rheumatology provided a synthetic data cohort of seronegative RA and PsA sine psoriasis, consistent with real data and privacy-safe,” the researchers concluded.
“Synthetic data improved classification accuracy. Key variables for intercepting PsA sine psoriasis included family history of PsO, axial/entheseal involvement and mismatch between pain measures and inflammatory markers.”
AI’s capacity to uncover new predictors of treatment response was highlighted in findings from the SPEED trial, where body mass index emerged as a stronger determinant of one-year psoriatic arthritis outcomes than treatment assignment itself.
The results, presented in a poster, emphasised the importance of integrating metabolic and lifestyle factors into individualised care strategies, the researchers said.
They analysed data from the SPEED trial, a multicentre, open-label study involving 192 patients with early PsA and poor prognostic factors. Participants were randomised to standard step-up conventional DMARD (csDMARD) therapy, combination csDMARD therapy, or early TNF inhibitor induction.
The study found that body mass index (BMI) was the single most influential baseline predictor of PASDAS outcomes at 48 weeks, outweighing even the type of treatment received. Patients with BMI under 25 kg/m² and lower baseline disease activity (PASDAS < 5.4) achieved the most favourable outcomes, regardless of therapy.
For patients with higher baseline disease activity (PASDAS ≥ 5.4) but BMI under 27 kg/m², early TNF inhibitor therapy led to better outcomes than standard or combination csDMARD treatment, highlighting the potential benefit of early biologic intervention.
Conversely, patients with higher BMI (≥ 27 kg/m²), longer disease duration (≥ 12 months), and polyarticular involvement, particularly those aged 51–74 years, experienced poorer outcomes, with predicted PASDAS scores reaching as high as 6.0.
“The predictive rules derived from this analysis offer clinicians valuable insights to support personalised care and encourage patients to prioritise weight management as a key component of their PsA treatment strategy,” the researchers concluded.
This content has been independently prepared by Rheumatology Republic with education funding from UCB Australia Proprietary Limited. The views expressed do not necessarily reflect the views of UCB Australia. AU-BK-2500275
