Novel Methods of Incorporating Time in Longitudinal Multivariate Analysis Reveals Hidden Associations With Disease Activity in Systemic Lupus Erythematosus

• Main Authors: Hieu T. Nim, Kathryn Connelly, Fabien B. Vincent, François Petitjean, Alberta Hoi, Rachel Koelmeyer, Sarah E. Boyd, Eric F. Morand
• Format: Article
• Language: English
• Published: Frontiers Media S.A. 2019-07-01
• Series: Frontiers in Immunology
• Subjects: systemic lupus erythematosus; biomarkers; clustering; longitudinal analysis; regression models
• Online Access: https://www.frontiersin.org/article/10.3389/fimmu.2019.01649/full

Objective: Systemic lupus erythematosus (SLE) is a multisystem autoimmune disease. SLE is characterized by high inter-patient variability, including fluctuations over time, a factor which most biomarker studies omit from consideration. We investigated relationships between disease activity and biomarker expression in SLE, using novel methods to control for time-dependent variability, in a proof-of-concept study to evaluate whether doing so revealed additional information.Methods: We measured 4 serum biomarkers (MIF, CCL2, CCL19, and CXCL10) and 13 routine clinical laboratory parameters, alongside disease activity measured by the SLE disease activity index-2k (SLEDAI-2k), collected longitudinally. We analyzed these data with unsupervised learning methods via ensemble clustering, incorporating temporal relationships using dynamic time warping for distance metric calculation.Results: Data from 843 visits in 110 patients (median age 47, 83% female) demonstrated highly heterogeneous time-dependent relationships between disease activity and biomarkers. Unbiased magnitude-based hierarchical clustering of biomarker expression levels isolated a patient subset (n = 9) with distinctively heterogeneous expression of the 17 biological parameters, and who had MIF, CCL2, CCL19, and CXCL10 levels that were higher and more strongly associated with disease activity, based on leave-one-out cross-validated regression analysis. In the remaining subgroup, a time-dependent regression model revealed significantly stronger predictive power of biomarkers for disease activity, compared to a time-agnostic regression model. Despite no significant difference in simple magnitude, using dynamic time warping analysis to align longitudinal profiles revealed a large subset (n = 69) with significantly stronger associations between biological parameters and disease activity. This subgroup had significantly lower flare rates, disease activity and damage scores, suggesting this clustering is clinically meaningful.Conclusions: These results suggest associations between biological parameters and disease activity in SLE exist in a multi-dimensional time-dependent pattern, with implications for the analysis of biomarkers in SLE often used to identify therapeutic targets. Novel methods to analyse high-dimensional data and control for time-dependent variability may have broad utility in the study complex relationships between clinical and biological parameters.