Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Hemispheres are identifiable based on connectivity; but handedness is not
Poster Session C, Saturday, September 13, 11:00 am - 12:30 pm, Field House
This poster is part of the Sandbox Series.
Trevor Day1, Peter Turkeltaub1, Elissa Newport1, Andrew DeMarco1; 1Georgetown University
Broadly speaking, the hemispheres of the brain are structurally and functionally very similar, but some functions are lateralized, notably including language processing. After a left hemisphere (LH) perinatal stroke, the right hemisphere (RH) can reorganize to assume the language functions of the LH (Newport et al., 2022), but this is no longer possible in older children or adults. There are also some differences in the organization of the brains of left and right-handers (Li et al., 2014), with reversed laterality for language being more common (but still the minority) in left-handers. We hypothesized that we could classify a hemisphere as a LH or RH and as belonging to a left- or right-hander, based on its intrinsic organization. While task activation is certainly suitable for this problem, resting-state (RS) is generalizable between studies. Finally, we also asked which brain regions are most important for classifying the hemisphere. We investigated whether a hemisphere can be identified as a LH or RH based on RS connectivity within that hemisphere. Using data from 935 healthy young adults from the Human Connectome Project (HCP), we created “hemiconnectomes:” partial connectomes created from ipsilateral connections. We asked if these hemiconnectomes are identifiable as belonging to an LH or RH and to a right- or left-handed individual. We tested three kinds of classifiers (Hannum et al., 2023) on RS data from HCP: a linear discriminant analysis (LDA), support vector classification, and a neural network. Each model was trained on pairwise correlations (n = 16,110) between all ipsilateral regions, and was asked to predict a four-way class, hemisphere ✕ handedness. Each connection receives a value reflecting its importance to the LDA classifier. To understand what regions were most important for classification, we computed weighted node centrality, a summary value of the strength of all connections to a given region. All three models perfectly identified hemisphere chirality in both handedness groups. However, none of the models were able to discriminate between handedness groups (Matthew’s correlation coefficient [MCC]: -.07 – .33). The small proportion of left-handers in HCP poses a class-imbalance problem. However, oversampling the left-handers did not improve the models (MCC: -.15 – .23). None of the model types performed reliably better than the others (post hoc Tukey adjusted p-values > .08). The networks with the highest node centrality (i.e. those that are most distinguishable between the hemispheres) are the cingulo-opercular (mean strength: 5.48), somatomotor (5.8), and language networks (5.28; Ji et al., 2019). Other high-strength regions are found throughout the temporal lobe. The next analytic step is to use hemiconnectomes calculated during the HCP language task, which may magnify differences between the handedness groups. The long-term goal of this project is application of the technique to the intact hemispheres of patients who suffered a perinatal or adult stroke, to measure how (and whether) their intact RH has reorganized to function more like a LH.
Topic Areas: Computational Approaches,