A MACHINE LEARNING APPROACH FOR PRECISION STRATIFICATION OF JUVENILE-ONSET SLE

Background: Juvenile-onset systemic lupus erythematosus (JSLE) is a complex and heterogeneous disease characterised by diagnosis and treatment delays. An unmet need exists to better characterise the immunological profile of JSLE patients and investigate its links with the disease trajectory over time. Objectives: A machine learning (ML) approach was applied to explore new diagnostic signatures for JSLE based on immune-phenotyping data and stratify patients by specific immune characteristics to investigate longitudinal clinical outcome. Methods: Immune-phenotyping of 28 T-cell, B-cell and myeloid-cell subsets in 67 age and sex-matched JSLE patients and 39 healthy controls (HCs) was performed by flow cytometry. A balanced random forest (BRF) ML predictive model was developed (10,000 decision trees). 10-fold cross validation, Sparse Partial Least Squares-Discriminant Analysis (sPLS-DA) and logistic regression was used to validate the model. Longitudinal clinical data were related to the immunological features identified by ML analysis. Results: The BRF-model discriminated JSLE patients from healthy controls with 91% prediction accuracy suggesting that JSLE patients could be distinguished from HCs with high confidence using immunological parameters. The top-ranked immunological features from the BRF-model were confirmed using sPLS-DA and logistic regression and included CD19+ unswitched memory B-cells, naïve B-cells, CD14 monocytes and total CD4 , CD8 and memory T-cell subsets. K-mean clustering was applied to stratify patients using the validated signature. Four groups were identified, each with a distinct immune and clinical profile. Notably, CD8 T-cell subsets were important in driving patient stratification while B-cell markers were similarly expressed across the JSLE cohort. JSLE patients with elevated effector memory CD8 T-cell frequencies had more persistently active disease over time, and this was associated with increased treatment burden and prevalence of lupus nephritis. Finally, network analysis identified specific clinical features associated with each of the top JSLE immune-signature variables. Conclusion: Using a combined ML approach, a distinct immune signature was identified that discriminated between JSLE patients and HCs and further stratified patients. This signature could have diagnostic and therapeutic implications. Further immunological association studies are warranted to develop data-driven personalised medicine approaches for JSLE. Acknowledgments: Lupus UK, Rosetrees Trust, Versus Arthritis Disclosure of Interests: George Robinson: None declared, Junjie Peng: None declared, Pierre Dönnes: None declared, Leda Coelewij: None declared, Meena Naja: None declared, Anna Radziszewska: None declared, Chris Wincup: None declared, Hannah Peckham: None declared, David Isenberg Consultant of: Study Investigator and Consultant to Genentech, Yiannis Ioannou: None declared, Ines Pineda Torra: None declared, Coziana Ciurtin Grant/research support from: Pfizer, Consultant of: Roche, Modern Biosciences, Elizabeth Jury: None declared Citation: Ann Rheum Dis, volume 79, supplement 1, year 2020, page 179Session: Novel diagnostic and therapeutic approaches in paediatric rheumatic diseases (Oral Presentations)