Unsupervised clustering of clinical and laboratory features of systemic lupus erythematosus: insights from a multicentre cohort

Objective This study explored the relationship between clinical phenotypes and immuno-molecular features of systemic lupus erythematosus (SLE) using unsupervised machine learning and multi-omics integration. Methods A multicentre cohort of 1065 SLE patients from five teaching hospitals was studied. Unsupervised clustering was performed using integrated clinical and laboratory data to identify distinct phenotypic clusters. Multi-omics profiling, including proteomics and transcriptomics, was conducted for each cluster, comparing with healthy controls. In addition, 899 patients were prospectively followed for over 1 year to assess therapeutic responses and remission outcomes. Results Unsupervised clustering identified five distinct clusters of SLE based on clinical phenotypes and immunological signatures, and an intrinsic association was found between each cluster and immune-inflammatory features. Cluster 1 (renal; 14.27%) had predominant renal involvement (87.5%) and high disease activity. Cluster 2 (mucocutaneous; 39.53%) exhibited mainly cutaneous manifestations (72.45%) with mild visceral involvement and the highest rates of achieving low disease activity state (46.42%). Cluster 3 (neuropsychiatric; 14.74%) had a high frequency of neuropsychiatric manifestations (96.18%) and anti-U1 ribonucleoprotein antibodies positivity (42.04%). Cluster 4 (haematologic; 22.63%) was characterised by haematologic involvement, predominantly cytopenia (95.85%), with frequent antiphospholipid antibody positivity and direct Coombs positivity (87.55%). Cluster 5 (cardiopulmonary; 8.83%) exhibited prominent cardiopulmonary involvement (97.87%) and serositis, with enrichment of inflammatory CD161 + regulatory T cells. Transcriptomic and proteomic analyses confirmed distinct molecular signatures and cluster-specific enrichment of biological pathways. Key molecule apolipoprotein A4 was validated using enzyme-linked immunosorbent assay. Regarding therapeutic outcomes at 1 year follow-up, Cluster 2 had the highest proportions achieving lupus low disease activity state (46.42%) and tapering glucocorticoids. Cyclophosphamide use was associated with greater clinical improvement (r=0.284; p=0.023) in Cluster 3, while in Cluster 5, rituximab use was associated with favourable responses (r=0.286; p=0.019). Conclusions We identified five SLE clusters with distinct clinical, immunological features and treatment outcomes, demonstrating an intrinsic link between clinical involvement and immune-inflammatory alterations.