Comprehensive examination of resting state fMRI connectomics yields new insights into brain function deficits in Gulf War illness after accounting for heterogeneity in brain impairment across the ill veteran population

Abstract

An estimated 200,000 veterans (up to 32% of those deployed) of the 1991 Gulf War (GW) suffer from GW illness (GWI), an incompletely understood chronic medical condition, characterized by multiple symptoms indicative of brain function deficits in various domains. Epidemiologic and animal studies have associated GWI with exposure to neurotoxic chemicals such as nerve agents, organophosphate pesticides and pyridostigmine bromide. One factor that hampers mechanistic investigations into GWI is that there is considerable heterogeneity in brain impairments across the ill GW veteran population. This could reflect the underlying heterogeneity in both exposure to neurotoxic substances, as well as genetic predisposition or resistance to neurotoxicity. Only one of the validated case definitions, the Haley GWI criteria addresses this heterogeneity. It does so by breaking down GWI into three main syndrome variants (GWS1, GWS2, and GWS3) based on factor analysis of symptoms presented by GWI veterans. Resting state fMRI (rsfMRI) is a uniquely useful brain imaging technique in that in a 10-min fMRI scan it can probe numerous brain function domains simultaneously. In this study, we employed a connectomics approach and machine learning on rsfMRI data from a cohort of GW veterans to extract neuroimaging biomarkers specific to each of the three Haley GWI syndromes. Our results revealed a number of new insights into brain function impairment specific to each syndrome group. The findings indicate that these deficits may by and large be driven by brain mechanisms. We also found that pooling the data of all three syndromes in GWI group, as is done by commonly employed case definitions of GWI resulted in failure to detect the fMRI signatures of a lot of these brain impairments.

Keywords: Biomarkers; Brain function networks; Connectomics; Functional connectivity; Gulf war illness; Heterogeneity; Neurotoxic; Ordinal support vector machine; Resting state fMRI.

Fig. 1

Schematic illustration for the Ordinal Support Vector Machine (Ordinal-SVM). The top panel shows the example of training a 4-class ordinal SVM classifier from N subject's connectivity matrices. The bottom panel shows the steps involved in predicting the ordinal classes of two new subjects (not included in the training dataset).

Fig. 2

Maps of Degree Centralities of Brainnetome nodes for (top) GWS1 vs VC; (middle) GWS2 vs VC; (bottom) GWS3 vs VC SVM classification boundaries. The maps show the number of edges connected to each Brainnetome node which are in the top 90th percentile of SVM_imp scores.

Fig. 3

(above) GWS1 vs VC 2-sample t-test on FC to left sensory thalamus node of the Brainnetome atlas, and (below) Nodewise SVMimp scores of the GWS1 vs VC SVM classification boundary of edges of the left sensory thalamus node. Only edges in the top 90th percentile of SVMimp scores are shown.

Fig. 4

(above) GWS1 vs VC 2-sample t-test on FC to left cerebellar lobule VIIIb node of the Brainnetome atlas, and (below) Nodewise SVM_imp scores of the GWS1 vs VC SVM classification boundary of edges of the left cerebellar lobule VIIIb node. Only edges in the top 90th percentile of SVM_imp scores are shown.

Fig. 5

(above) GWS2 vs VC 2-sample t-test on FC to left perirhinal and entorhinal cortex node of the Brainnetome atlas, and (below) Nodewise SVM_imp scores of the GWS1 vs VC SVM classification boundary of edges of the left perirhinal and entorhinal cortex node. Only edges in the top 90th percentile of SVM_imp scores are shown.

Fig. 6

(above) GWS2 vs VC 2-sample t-test on FC to right cerebellar lobule VIIIB node of the Brainnetome atlas, and (below) Nodewise SVM_imp scores of the GWS1 vs VC SVM classification boundary of edges of the right cerebellar lobule VIIIb node. Only edges in the top 90th percentile of SVM_imp scores are shown.

Fig. 7

(above) GWS3 vs VC 2-sample t-test on FC to left ventral caudate node of the Brainnetome atlas, and (below) Nodewise SVM_imp scores of the GWS1 vs VC SVM classification boundary of edges of the left ventral caudate node. Only edges in the top 90th percentile of SVM_imp scores are shown.

Fig. 8

(above) GWS3 vs VC 2-sample t-test on FC to left cerebellar lobule VIIIa node of the Brainnetome atlas, and (below) Nodewise SVM_imp scores of the GWS1 vs VC SVM classification boundary of edges of the left cerebellar lobule VIIIa node. Only edges in the top 90th percentile of SVM_imp scores are shown.