Shape features of white matter tracts associated with post-surgical speech production outcomes

• Impairments in speech production occur commonly after surgery for brain tumours. • Tumours have high spatial heterogeneity and can displace or deform adjacent tissue. • Tumours are problematic for lesion-symptom mapping and disconnection approaches. • Individualised tractography shape features better predict post-surgical outcomes. Impairments in speech production occur at a high rate after surgical resection of primary brain tumours. Due to their spatial heterogeneity, tumours can be problematic for voxel-wise lesion-symptom mapping. In addition, lesion maps of resection cavities may not capture displaced/deformed tracts, limiting the accuracy of white matter disconnection approaches based on population average atlas templates. Here we explore whether shape features extracted from individualised diffusion MRI tractography such as volume, curvature, surface area, irregularity, and span of six language-related white matter tracts are able to predict long-term speech production outcomes in survivors of left hemisphere primary brain tumours: The arcuate (AF), uncinate (UF), inferior longitudinal (ILF), inferior frontal occipital (IFOF), superior longitudinal (SLF) fasciculi, and frontal aslant tract (FAT). Best fit regression models selected via cross-validation and both univariate and multivariate lesion-symptom mapping analyses including clinical and demographic variables were used to compare the predictive capability of tract shape features, atlas disconnection estimates, and lesion location. Lesion-symptom mapping failed to reveal any significant relationships with speech production outcomes. A model comprising atlas-based disconnection estimates from the UF, SLF, and FAT was able to significantly predict long-term speech production outcomes ( R 2 = 0.14). However, a model comprising shape features of the FAT, ventral (III) component of the SLF, and IFOF explained considerably more variance ( R 2 = 0.37). Finally, we show that, in this dataset, shape features uniquely predict 26% variance that is not accounted for by disconnection estimates, demonstrating their relevance for predicting speech production outcomes in post-surgical patients.