Machine Learning Classification of Neuropsychiatric Systemic Lupus Erythematosus patients using resting-state fMRI functional connectivity

Abstract

In this study we explored the robustness of machine learning algorithms for the classification of Neuropsychiatric systemic lupus erythematosus (NPSLE) patients and healthy controls using resting-state fMRI functional connectivity matrices. NPSLE, which is driven by systemic autoimmune inflammation in the context of lupus, involves a wide range of focal and diffuse central and peripheral nervous system symptoms and poses significant diagnostic challenges. Machine learning applications on clinical data may enhance the existing workflow for NPSLE classification as there is no established method of applying neuroimaging data to the diagnosis of NPSLE. Feature selection methods were applied prior to the classification process in order to perform the classification process on a lower dimension feature space. The Connectivity Matrix used consisted of pairwise regional functional associations of the fMRI signals (ROI to ROI correlations) within each of three predetermined brain networks in 41 NPSLE patients and 31 healthy control subjects. Support Vector Machines (SVM) was utilized in the final model. Results were evaluated using a nested cross validation methodology to prevent overfitting, and enhance generalization. Regions of Interest (ROI's) that contributed most in the final model were: Right Inferior Temporal, Thalamus, Left Angular Gyrus, Right Precuneus, Left Primary Motor Cortex, SMA, Left and Right Primary Motor Cortex. With a final F1 score of up to 77%, the results demonstrate the potential for the future implementation of similar methods in the diagnosis of NPSLE.