Background: Generative artificial intelligence (AI) may enable privacy-preserving cohort enrichment by generating synthetic patient profiles that reproduce multivariable distributions and correlations. To provide the first proof-of-concept study in rheumatology we chose the highly relevant issue of distinguishing seronegative rheumatoid arthritis (RA) from psoriatic arthritis without psoriasis (PsA sine psoriasis). In fact, this remains challenging with diagnoses often drifting over time. Since several newer drugs are effective only in either condition, early accurate classification is even more clinically relevant.

Objectives: To validate synthetic data generation in inflammatory arthritis and assess whether synthetic augmentation improves machine-learning discrimination and explainability of seronegative RA versus PsA sine psoriasis and data-driven patient phenotyping.

Methods: We analyzed a retrospective cross-sectional cohort of 200 adults, 95 with seronegative RA and 105 PsA sine psoriasis (70% females; median age at onset 49 years (IQR 39-58)). Variables collected at first specialist evaluation included demographics, smoking, BMI, family history of psoriasis, comorbidities, baseline phenotype (mono/oligo/polyarthritis, symmetrical or asymmetrical arthritis, axial involvement, enthesitis, dactylitis, nail disease, inflammatory bowel disease, interstitial lung disease) and baseline activity indices (CRP, tender (TJC) and swollen joint counts (SJC), pain VAS, physician global assessment). Treatment history was abstracted (csDMARDs and b/tsDMARDs grouped by mechanism), including number of mechanisms tried and the last effective category (csDMARD, b/tsDMARDs for RA-only, b/tsDMARDs for PsA-only, shared b/tsDMARDs).

Synthetic profiles were generated using a conditional Wasserstein tabular GAN with gradient penalty and validated with SAFE (D’Amico, 2023) using Clinical Synthetic Fidelity (CSF) and distance-to-closest-record and nearest-neighbour distance ratio (NNDR). Clinical utility was evaluated with (a) Random Forest and Multilayer Perceptron (MLP) classifiers using 3-fold cross-validation and SHAP explanations, and (b) with HDBSCAN clustering, with diagnoses and treatments were examined post hoc.

Results: In the real-originator seronegative RA group, symmetrical arthritis was common (88%) with higher joint counts (median TJC 8 [IQR 4-12], SJC 2 [IQR 0-6]) and CRP 0.9 (IQR 0.1-2.5) mg/dL. In the real PsA sine psoriasis group, peripheral arthritis was present in 86% (symmetrical 58%) with lower joint counts (median TJC 4 [IQR 2-9], SJC 1 [IQR 0-3]) and CRP 0.8 (IQR 0.1-2.5) mg/dL; axial involvement was reported in 33% and enthesitis in 18%. Fibromyalgia was more prevalent in PsA sine psoriasis than in seronegative RA (26% vs 8%).

A synthetic cohort of 200 profiles closely recapitulated the distributions and correlation structure (CSF 92%) of the real cohort without generating any exact duplicate and preserving privacy (NNDR 0.73). Synthetic patients maintained similar feature distribution and correlations observed in the originator cohort.

Classifiers were first trained on 160 real patients and benchmarked against 40 unseen real patients. On real data only, the classification performance was 86% (±4.2) for Random Forest and 82% (±3.2) for MLP. After integrating synthetic patients in the train set, the performance of the classifiers increased to 87% (±3.1) for Random Forest and 84% (±3.2) for MLP. Synthetic augmentation also shifted the most influential features identified by SHAP: in models trained on real data only, symmetrical peripheral arthritis and CRP mainly drove RA classification, while family history of psoriasis, axial involvement, and fibromyalgia drove PsA. After adding synthetic data, older age and polyarthritis pattern emerged among strong RA drivers together with a higher inflammatory burden (CRP and SJC). Instead, PsA classification was more strongly driven by asymmetric joint involvement and a pain-inflammation mismatch (more fibromyalgia and high TJC but low SJC and CRP).

Clustering on real data identified five clinical groups: a RA-enriched cluster (predominantly female, low comorbidity, symmetric polyarthritis), two PsA-enriched clusters (one with axial/entheseal involvement and IBD; the other male-predominant with asymmetric arthritis, dactylitis, and metabolic comorbidities), and two mixed clusters combining both features suggestive of PsA and RA. Treatment associations were consistent with the clinical phenotypes: the RA-enriched cluster responded mostly to csDMARDs, the PsA-enriched clusters mostly to PsA-only b/tsDMARDs.

After adding synthetic patients, clustering led to a clearer separation of the clinical profiles. Two RA-predominant clusters were identified: one predominantly female with low comorbidity burden, the other with predominant oligoarthritis. PsA-like clusters became more distinct, encompassing a group with mono-oligoarticular involvement, and another characterized by axial disease, enthesitis and IBD. Moreover, after integrating synthetic data, treatment patterns aligned more consistently with phenotype clusters. High-comorbidity clusters showed the greatest therapeutic complexity (multiple mechanisms tried and higher prevalence of treatment failure).

Conclusions: Our first proof-of-concept study in Rheumatology demonstrates that synthetic data generation is feasible, privacy-preserving, and clinically coherent for inflammatory arthritis. Leveraging different machine learning architectures, synthetic augmentation yields modest gains in discriminating between seronegative RA versus PsA sine psoriasis. Integration of synthetic profiles provides data augmentation and strengthens models’ explainability, identifying the relationship between inflammatory and pain burden as a major driver of RA versus PsA classification. Synthetic data also refined phenotype stratification into clinically actionable clusters, with more coherent therapeutic needs and outcomes.

REFERENCES: NIL.

Acknowledgments: NIL.

Disclosure of Interests: None declared.