Beyond ‘Dr Google’: Performance of Large Language Models in Patient Counselling for Connective Tissue Diseases
Abstract format and assignment number: PARE Abstract – OP013-PARE
Date: Wednesday, 3 June 2026
Presenting author: P. Kremer (Germany)
This PARE session compared Claude 4.0 Sonnet, ChatGPT-5, Gemini 2.5 Pro, and Google Search across 20 FAQs per connective tissue disease (SLE, IIM, SjD, SSc), rated by 20 patients and 5 rheumatologists. All LLMs were medically accurate and rated positively. Gemini 2.5 Pro ranked first most often (59% patients, 63% physicians), while Google Search was rated worst (55% patients, 57% physicians). LLMs outperformed Google Search on clarity and empathy, supporting their complementary role in patient education with appropriate safeguards.
Lighten the Load: Artificial Intelligence Reveals Body Mass Index Outranks Treatment in One-Year Psoriatic Arthritis Outcomes—Findings from the SPEED Trial
Abstract format and assignment number: Oral – OP0187
Date: Thursday, 4 June 2026
Presenting author: A. Garaïman (United Kingdom)
The SPEED trial (192 early PsA patients) was analyzed using RPART machine learning. BMI emerged as the strongest predictor of PASDAS at 48 weeks, ranking ahead of treatment. Patients with BMI below 25 kg/m² and low baseline disease activity achieved the best outcomes (predicted PASDAS 2.1), regardless of therapy. In those with higher BMI and polyarticular disease, early TNF inhibitor therapy outperformed csDMARDs. Findings underscore weight management as a key—and modifiable—component of PsA care.
Sharp.AI: Longitudinal AI-Based Evaluation of Radiographic Progression in Psoriatic Arthritis
Abstract format and assignment number: Poster – POS0668
Date: Thursday, 4 June 2026
Presenting author: Z. Gao (United States of America)
Sharp.AI, trained on radiographs from 4,150 PsA and RA patients across six clinical trials, achieved patient-level ICC 0.97 against expert panel vdHS scores—outperforming human inter-rater agreement (ICC 0.93). For vdHS change at week 24, ICC was 0.60 (vs. 0.59 for raters); at week 52, ICC reached 0.71 (vs. 0.64). Attention maps confirmed the model focused on clinically relevant joint regions, validating AI-based longitudinal radiographic scoring as a scalable, objective alternative to manual expert assessment.
Artificial Intelligence for Identifying and Grading Microcrystalline Deposits in Knee Ultrasound Images: A Semantic Segmentation Approach. Results of an OMERACT Ultrasound Working Group Project
Abstract format and assignment number: Poster – POS0782
Date: Thursday, 4 June 2026
Presenting author: M. D. Dal Fabbro (Italy)
A LinkNet/ResNet34 semantic segmentation model was trained on 279 expert-labelled knee ultrasound images from 14 international OMERACT centres. Mean IoU was 0.86 for background, 0.60 for meniscus, and 0.37 for CPPD deposits. Using area ratio thresholds, grade classification accuracy on predicted masks reached 64%. While current performance did not yet exceed simpler approaches—likely reflecting limited dataset size—the framework establishes a foundation for AI-driven CPPD grading as larger, more varied training sets become available.
Patients Improve, Patterns Persist: Temporal Stability of RA Subsets
Abstract format and assignment number: Poster – POS1271
Date: Saturday, 6 June 2026
Presenting author: T. D. Maarseveen (Netherlands)
Three DMARD-naive early RA cohorts (BeSt, NORD-STAR Swedish/Finnish, RZWN; over 1,400 patients) were analyzed. Despite treatment-driven joint count reductions, patients with similar baseline joint involvement patterns (JIPs) clustered together throughout the year (BeSt silhouette score 0.25, NORD-STAR 0.30, RZWN 0.50). The JIP-poly subgroup was the sole exception in trial populations. Findings confirmed JIPs represent distinct, persistent RA subtypes rather than transient disease stages, supporting their use as a foundation for precision medicine in early RA.
Dr Piotr Kuszmiersz MD PhD (Poland)
Country: Poland
Piotr is a PhD in Medical Sciences and an Internal Medicine Specialist, currently undertaking his rheumatology training at the Department of Rheumatology and Immunology, Jagiellonian University Medical College in Kraków, Poland. His research interests focus on Patient-Reported Outcome Measures, epidemiology, and the application of AI in medicine. Piotr is a member of the EMEUNET Country Liaison Sub-Committee.