Overlapping connectivity patterns during semantic processing of abstract and concrete words revealed with multivariate Granger Causality analysis

Abstract, unlike concrete, nouns refer to notions beyond our perception. Even though there is no consensus among linguists as to what exactly constitutes a concrete or abstract word, neuroscientists found clear evidence of a “concreteness” effect. This can, for instance, be seen in patients with language impairments due to brain injury or developmental disorder who are capable of perceiving one category better than another. Even though the results are inconclusive, neuroimaging studies on healthy subjects also provide a spatial and temporal account of differences in the processing of abstract versus concrete words. A description of the neural pathways during abstract word reading, the manner in which the connectivity patterns develop over the different stages of lexical and semantic processing compared to that of concrete word processing are still debated. We conducted a high-density EEG study on 24 healthy young volunteers using an implicit categorization task. From this, we obtained high spatio-temporal resolution data and, by means of source reconstruction, reduced the effect of signal mixing observed on scalp level. A multivariate, time-varying and directional method of analyzing connectivity based on the concept of Granger Causality (Partial Directed Coherence) revealed a dynamic network that transfers information from the right superior occipital lobe along the ventral and dorsal streams towards the anterior temporal and orbitofrontal lobes of both hemispheres. Some regions along these pathways appear to be primarily involved in either receiving or sending information. A clear difference in information transfer of abstract and concrete words was observed during the time window of semantic processing, specifically for information transferred towards the left anterior temporal lobe. Further exploratory analysis confirmed a generally stronger connectivity pattern for processing concrete words. We believe our study could guide future research towards a more refined theory of abstract word processing in the brain.

b) The whiteness test which necessitates that the residuals be uncorrelated white noise. For our study, we used the Li-McLeod Portmanteau test modified for multi-variate models 2 . c) The stability test which requires the eigenvalues of the autoregressive parameters to be smaller than 1. d) The percent consistency which is a comparison of the correlation vector of the real and reconstructed data using the autoregressive parameters 3 .
After performing a parameter exploration, we settled for a model order of 14 (corresponding to a time lag of 70ms) and an adaptation coefficient of 0.01 which passed the whiteness test and is also in line with simulation studies on the multi-trial General Linear Kalman filter 4,5 .
Some studies further validate the model order selection by comparing the power spectrum of the signal using a non-parametric Welch and parametric Burg method 6 . In our parameter exploration, we realized that the criteria described above do not converge to the same optimal parameters. For example, it has been previously mentioned that increasing the adaptation coefficient increases the speed of adaptation at the cost of losing smoothness of the estimates 4 . We also observed that a high adaptation coefficient increases the correlation of the residuals and decreases stability. However, a lower adaptation coefficient decreased the percent consistency. The parameter exploration of a sample subject is shown in figure S2. As can be seen in figure S2, there is always a trade-off between the different criteria defined in the literature. However, it should also be noted that these criteria were designed for purposes other than Granger connectivity. In our case, the most important criterion for Granger connectivity is consistency. Therefore, we performed a bootstrapping analysis of the Granger causality (described in section Statistical analysis of Granger connectivity).

Appendix C. Partial Directed Coherence results
The results for abstract trials are shown below.

Figure S3
Within-subject cluster-based non-parametric permutation test with 10,000 iterations for abstract trials. All shown results are significant for p<0.0001. Figure S4 Within-subject cluster-based non-parametric permutation test 10,000 iterations for concrete trials. All shown results are significant for p<0.0001.

Appendix D. Exploratory analysis of statistically significant connections
An exploratory analysis of abstract versus concrete words was performed for every 100ms and for all connections with an effect size above 0.8. The p value was set to 0.01 (uncorrected). We did not correct for multiple comparisons over time and connection pairs. However, similar to the main results we only considered results which were significant for a frequency band of minimum 3Hz. Additionally, results were only considered between 5 and 25 Hz, as we were apprehensive of edge artifacts outside this frequency range.

Appendix E. Inter-subject variability of connection pairs
In order to gauge the effect of inter-subject variability, we separated all time-frequency data points per subject into significant and non-significant connection pairs. The histogram of the standard deviation and the mean over subjects are shown in figure S5. As can be seen, the standard deviation of non-significant connections was much smaller than that of the significant connections. This suggests that when a connection pair is not significant, it is fairly consistent between subjects. However, the significant connections vary across subjects.