Quantitative Identification of Functional Connectivity Disturbances in Neuropsychiatric Lupus Based on Resting-State fMRI: A Robust Machine Learning Approach

Abstract

Neuropsychiatric systemic lupus erythematosus (NPSLE) is an autoimmune entity comprised of heterogenous syndromes affecting both the peripheral and central nervous system. Research on the pathophysiological substrate of NPSLE manifestations, including functional neuroimaging studies, is extremely limited. The present study examined person-specific patterns of whole-brain functional connectivity in NPSLE patients (n = 44) and age-matched healthy control participants (n = 39). Static functional connectivity graphs were calculated comprised of connection strengths between 90 brain regions. These connections were subsequently filtered through rigorous surrogate analysis, a technique borrowed from physics, novel to neuroimaging. Next, global as well as nodal network metrics were estimated for each individual functional brain network and were input to a robust machine learning algorithm consisting of a random forest feature selection and nested cross-validation strategy. The proposed pipeline is data-driven in its entirety, and several tests were performed in order to ensure model robustness. The best-fitting model utilizing nodal graph metrics for 11 brain regions was associated with 73.5% accuracy (74.5% sensitivity and 73% specificity) in discriminating NPSLE from healthy individuals with adequate statistical power. Closer inspection of graph metric values suggested an increased role within the functional brain network in NSPLE (indicated by higher nodal degree, local efficiency, betweenness centrality, or eigenvalue efficiency) as compared to healthy controls for seven brain regions and a reduced role for four areas. These findings corroborate earlier work regarding hemodynamic disturbances in these brain regions in NPSLE. The validity of the results is further supported by significant associations of certain selected graph metrics with accumulated organ damage incurred by lupus, with visuomotor performance and mental flexibility scores obtained independently from NPSLE patients.

Chronic Mild Traumatic Brain Injury: Aberrant Static and Dynamic Connectomic Features Identified Through Machine Learning Model Fusion

“…Initial data preparation steps are in line with several rs-fMRI studies (Luppi et al, 2019 ; Luppi, Carhart-Harris et al, 2021 ), including previous work of our team on data from the same MRI system (Antypa et al, 2021 ; Kavroulakis et al, 2021 ; Pentari et al, 2022 ; Simos et al, 2019 ; Simos et al, 2020 ). Firstly, the first three volumes were discarded to allow for magnetization effects to stabilize.…”

“…In terms of feature selection and method validation, a consensus feature importance ranking-based, nested cross-validation methodology was implemented similar to Parvandeh et al ( 2020 ), Zhong, Chalise, and He ( 2020 ), and our previous work (Simos et al, 2019 ; 2020 ). This choice was made to avoid overfitting and ensure a robust final model with conservative results and highly representative features for the comparison of interest.…”

Machine Learning for Diagnosis of Systemic Lupus Erythematosus: A Systematic Review and Meta-Analysis

“…Eleven hundred and sixty-six (1166) ineligible articles were excluded after browsing the abstracts and titles, and full texts of the remaining sixty (60) articles were read. Finally, a total of eighteen (18) studies were included, in which ten (10) [15][16][17][18][19][20][21][22][23][24] focused on SLE and the remaining eight (8) [25][26][27][28][29][30][31][32] on NPSLE. Te study selection process is shown in Figure 1, and the characteristics of included studies are shown in Tables 1 (for SLE) and 2 (for NPSLE).…”

Communities and Cliques in Functional Brain Network Using Multiscale Consensus Approach

“…The sparsified weighted networks are then either used as-is or binarized to give unweighted networks. Several edge-filtering procedures are effective in FCN studies [3]; however, the challenge lies in finding an optimal cutoff or threshold for filtering without losing salient information. The solutions widely use datadriven preprocessing procedures, and are dependent on the intended application.…”