Presentation

Search Abstracts | Symposia | Slide Sessions | Poster Sessions

Using Unsupervised Dimensionality Reduction to Identify Lesion Patterns Predictive of Post-Stroke Aphasia Severity

Poster C32 in Poster Session C, Friday, October 25, 4:30 - 6:00 pm, Great Hall 4

Emerson Kropp¹, Maria Varkanitsa¹, Nicole Carvalho¹, Isaac Falconer¹, Anne Billot², Swathi Kiran¹; ¹Boston University, ²Harvard University

Lesion volume and location are known to impact aphasia severity in stroke patients, particularly in left hemisphere regions (Døli et al., 2021). To investigate these impacts, we used non-negative matrix factorization (NMF), a data-driven approach that identifies representative patterns of lesioned brain regions (Bonkhoff et al., 2021; Kernbach et al., 2023) and explored associations with aphasia severity. Lesions were segmented using T1 structural MRIs for 98 left hemisphere stroke patients with aphasia. The Western Aphasia Battery - Revised Aphasia Quotient (WAB-R AQ) was used to quantify aphasia severity (mean age=61.9, mean WAB-R AQ=74.7). Left hemisphere percent spared tissue was calculated in 83 AAL3 grey matter regions and 36 white matter tracts from BCBToolkit (Rolls et al., 2020; Foulon et al., 2018). NMF is a dimensionality reduction method which, unlike principal component analysis, generates interpretable components by creating a parts-based representation of the information. By applying NMF to the region-wise tissue spared data, we reduced the dataset into a small number of lesion atoms which represent prototypical patterns of regions which tend to be damaged together across patients. The number of atoms was picked to minimize deviation between the original and reconstructed datasets (Kim & Tidor, 2003). Percent tissue spared was multiplied by region volumes to identify spared volume, then weighted based on relative region contributions to each atom and normalized (Thakallapalli et al., 2021); these weighted averages were used as aggregate values of each patient’s percent spared tissue within each spatially distributed lesion atom. Atom-wise spared tissue data was then used in linear regression models predicting WAB-R AQ (Marchina et al., 2011). Models were adjusted for gender, age, months post onset, education, and total lesion volume. 4 lesion atoms were identified. Atom 1 contained subcortical/posterior regions and was least associated with WAB-R AQ (Atom 1 partial-R2=.050, p<.05). Atom 2 contained mainly fronto-parietal regions such as the pre/postcentral gyri and superior longitudinal fasciculus (Atom 2 partial-R2=.086, p<.01). Atoms 3 and 4 both bordered the Sylvian fissure with overlap including the insula, Rolandic operculum, supramarginal, and Heschl’s gyri. Atom 3 had more contribution from frontal lobe regions, while atom 4 was localized mainly to the temporal lobe (Atom 3 partial-R2=.12, p<.001; Atom 4 partial-R2=.15, p<.0001). Atom 4’s model was most associated with WAB-R AQ overall (adjusted-R2=.30, p=1.7e-6). When comparing region-wise contributions, the highest increases in contribution to atom 4 versus atom 3 were in the superior temporal, middle temporal, and Heschl’s gyri, the Rolandic operculum, and the posterior arcuate fasciculus. Atoms 3 and 4 were the most associated with aphasia severity, and each contained classical language regions in the left MCA territory. However, the improved performance of atom 4 compared to 3 conveys the importance of temporal lobe white and grey matter over other regions in predicting aphasia severity. Each atom is a representative pattern of damage, suggesting that atom 4 is a natural lesion pattern particularly relevant to aphasia. Better understanding of this pattern’s etiology may provide insight into the mechanisms of language dysfunction in stroke.

Topic Areas: Disorders: Acquired,