Background/Purpose: Systemic Lupus Erythematosus (SLE) is a chronic auto-immune condition characterized by systemic inflammation that can exacerbate (flare) unpredictably, causing widespread organ damage.  Identification of SLE flares, particularly at an early stage, remains a challenge for rheumatologists because many flare symptoms mimic those of other conditions but is critical in order to resolve systemic inflammation and spare patients unnecessary treatment if their symptoms are not SLE-related.

Methods: We use a machine-learning approach to classify patient-reported flare status from proteomic measurements in serum from a cohort of 144 individuals with SLE. Proteomic measurements collected from serum samples were inputed to a logistic regression model. The cohort was split 70-30% into training and test sets; 5 fold cross validation was performed within the training set to select hyperparameters. Performance in the test set is reported as area under the receiver operating characterisic curve as an average from 100 test-training split iterations.

Results: We find improvement in the performance of a logistic regression model of patient-reported flare status when the substrate feature pool is reduced from the full proteomic panel of 900 proteins to a 6-protein panel of cytokines selected through curation of peer-reviewed biomedical literature. Use of this prior knowledge protein panel as input to the learning algorithm boosts performance from AUC=0.66 using the full panel to AUC=0.83 using the 6-feature selected panel.

Conclusion: Our results point to a critical role for prior experimental knowledge of molecular disease drivers in creating accurate mathematical, molecular models with diagnostic potential.

« Back to 2019 ACR/ARP Annual Meeting

ACR Meeting Abstracts - https://acrabstracts.org/abstract/prior-knowledge-feature-reduction-improves-performance-in-a-machine-learning-model-of-systemic-lupus-erythematosus-flare-status-using-serum-proteomics/