A Connectome-Based Predictive Modelling and Graph-Theoretical Study of Language Performance in Later Adulthood

Language performance varies substantially across individuals in later adulthood, including vocabulary comprehension and single-word reading abilities. Although previous studies have shown that individual differences in language performance can be predicted from resting-state functional connectivity (rs-FC) in younger adults, it remains unclear whether such predictive relationships extend to later adulthood. It is also unclear how a functionally defined language network (LANG) is represented within these predictive connectivity patterns. Furthermore, whether the topology of the LANG is associated with language performance, and whether these associations are moderated by key demographic factors such as age and education, remains unclear. The present study used resting-state fMRI and behavioural data of 445 adults aged 40–86 years from the Human Connectome Project Aging dataset. Language performance was assessed using two NIH Toolbox measures: the Picture Vocabulary Test (TPVT), indexing receptive vocabulary comprehension, and the Oral Reading Recognition Test (READ), indexing single-word reading recognition. The LANG was constructed by overlapping a functionally-defined language rs-FC map with the Schaefer 400-parcel atlas, yielding 29 language parcels. Whole-brain connectome-based predictive modelling (CPM) examined whether rs-FC was able to predict the language performance, in other words, whether rs-FC contained information relevant to individual differences in language performance, after adjusting for age, education, and mean framewise displacement (FD). Stable model-selected edges were ranked according to their descriptive contribution to prediction and characterised by their involvement of different large-scale brain networks. In parallel, graph-theoretical general linear models (GLMs) examined two topological properties of LANG: global efficiency, reflecting within-network integration, and node-wise participation coefficient (PC), reflecting cross-network connectivity. Age and education were examined as potential moderators of the associations between LANG topology and language performance, with mean FD also included as a covariate. The CPM results showed that rs-FC predicted both TPVT and READ with statistically reliable but modest predictive performance. Model-selected edges were distributed across multiple large-scale networks, indicating that language performance in later adulthood was supported by distributed connectivity patterns rather than by the language network alone. Graph-theoretical analyses further suggested that associations between LANG topology and performance were mainly expressed through moderation effects rather than main effects. Education was the more prominent moderator, particularly for TPVT, whereas age-related moderation was observed primarily in node-wise analyses across TPVT and READ. These findings together highlight the distributed pattern of rs-FC for characterising individual differences in later-life vocabulary comprehension and word recognition, while suggesting that demographic factors, especially education, should be considered when interpreting brain–language associations in older adults.