Which extra-renal flare is ‘difficult to treat’ in systemic lupus erythematosus? A one-year longitudinal study comparing traditional and machine learning approaches

Abstract

Abstract Objectives To describe phenotypes and outcomes of extra-renal flares in SLE, to identify clusters of extra-renal flares based on baseline features, and to develop a machine learning (ML) tool capable of predicting ‘difficult to treat’ (D2T) flares. Methods Extra-renal flares that occurred in our cohort over the last five years with at least one year of follow-up were included. Baseline clinical variables were described and flares assigned to clusters. Attainment of remission and low disease activity state (LLDAS) at 12 months were compared. Flares were then considered ‘D2T’ in case of non-attainment of LLDAS at 6 and 12 months. Baseline features were used to train a ML model able to predict future D2T-flares, at admission. Traditional approaches were then compared with informatic techniques. Results Among 420 SLE patients of the cohort, 114 flares occurred between 2015 and 2021; 79 extra-renal flares, predominantly mucocutaneous (24.1%) and musculoskeletal (45.6%), were considered. After 12 months, 79.4% and 49.4% were in LLDAS and in remission, respectively, while 17 flares were classified as D2T (21.5%); D2T flares received a higher cumulative and daily dose of glucocorticoids. Among the clusters, cluster ‘D’ (mild-moderate flares with mucocutaneous manifestations in patients with history of skin involvement) was associated with the lowest rate of remission. Among clinical data, not being on LLDAS at 3 months was the unique independent predictor of D2T flares. Conclusions Our clusterization well separates extra-renal flares according to their baseline features and may propose a new identification standard. D2T flares, especially refractory skin manifestations, are frequent in SLE and represent an unmet need in the management of the disease as they are associated with higher glucocorticoid (GC) dosage and risk of damage accrual. Our ML model could help in the early identification of D2T flares, flagging them to elevate the attention threshold at admission.