Word finding difficulty is a prevalent cognitive symptom in multiple sclerosis (MS). Word finding relies on retrieving concepts and word forms from the long-term store. Neuropsychological assessment of word finding difficulty in persons with MS (pwMS) is typically characterized by semantic errors and decreased speed in naming tests, along with decreased semantic verbal fluency scores. Despite this, there is significant heterogeneity in the detection of verbal fluency deficits across studies in the MS literature. This may be partially due to disease-related heterogeneity and/or low sensitivity of commonly used scoring approaches. We investigate the latter in the present study. Semantic network analysis, derived from graph theory, provides a fine-grained approach to understanding semantic retrieval by utilizing information about the co-occurrence of words produced on semantic verbal fluency tasks. Analysis results in a graphical quantification of the conceptual-lexical store. A preliminary study found that semantic networks from Spanish-speaking pwMS had fewer associative connections and more central connective pathways, which if affected, may lead parts of the network to become inaccessible for retrieval. However, their investigation was limited in the generalizability of their findings, as they excluded pwMS who have cognitive impairment (CI), which represents a significant proportion of pwMS. We sought to investigate network differences in an English-speaking MS sample, without exclusion based on CI, using widely-used metrics of micro-, meso-, and macroscopic structure. We hypothesize the MS network will be less efficiently organized, thus characterized by higher average shortest path length (ASPL), lower clustering coefficient (CC) and lower modularity (Q).

53 persons with MS and 44 neurologically healthy controls (HC) were recruited as a part of an ongoing study (NMSS RG-1907-34364 & RG-1901-33304). As a part of a larger battery, participants were administered the semantic verbal fluency subtest of the Controlled Oral Word Association Test. Responses were analyzed using a network-analysis R suite.

The MS and HC networks were characterized by having similar average shortest path lengths (ASPL MS =2.466, ASPL HC=2.463, F(1,1997)=0.281, p=0.596), indicating they require similar numbers of edges to be traversed to reach other nodes in the network. This suggests similar efficiency of information transfer. Clustering coefficient was not significantly different between the MS and HC networks (CC MS = 0.742, CC HC =0.742, F(1,1997)=0.10, p=0.919), suggesting similar local interconnectivity. The MS network had significantly lower modularity compared to the HC network (Q MS =0.497, Q HC = 0.502, F(1,1997)=16.678, p<0.001). This means that sub-communities of the network were less segregated into densely connected sub-graphs.

Contrary to expectation, ASPL and CC were not significantly different between groups. The absence of finding lower CC was consistent with prior findings. Consistent with our hypothesis, the MS network had lower modularity. This may suggest that pwMS were unable to use categorical clustering to aid in retrieval from the lexicon. Specifically, low modularity coupled with similar CC may suggest the structure of the MS lexicon is characterized by intact clustering on a microscopic scale but less strong organization into distinct clusters on a larger scale.