Slide Slam H5
A comparison of observed and simulated disconnection measures in chronic post-stroke aphasia.
Ajay Halai1, Matthew Lambon Ralph1; 1MRC Cognition and Brain Sciences Unit, University of Cambridge, UK
There is a long-standing history of mapping behavioural deficits to brain damage in stroke patients (i.e., inferior frontal regions for speech production and posterior temporal regions for speech comprehension). Research studies typically focus on grey matter but we know that white matter (dis-)connections are important (Catani et al., 2005). In practice, collecting diffusion weighted/tensor imaging (DWI/DTI) data in vulnerable populations is challenging as: a) they not essential for clinical assessment; b) they require expert analytics and c) are time consuming/expensive. This has led to multiple algorithms being developed to infer disconnections given a particular lesion profile. A number of studies within the stroke literature have utilised inferred disconnections to build prediction/prognostic models (e.g. Pustina et al., 2018; Hope et al., 2018) but have had mixed outcomes. In particular, a recent study has compared predictions using lesion location with inferred structural and functional disconnections (Salvalaggio et al., 2020). Inferred disconnections are typically obtained using the lesion location (as a seed) in healthy younger adults, which assumes that: a) changes due to normal aging are negligible; and b) spared regions in stroke patients are ‘healthy’ (specifically the contralateral hemisphere). To date, there has been no formal comparison of the observed disconnections in-vivo with a range of inferred disconnections; therefore, this study set out to address this question. In addition, we also used lesion location and each disconnection method as features in a 5-fold cross validation prediction analysis. We obtained in-vivo diffusion data using the same sequence parameters from left-sided chronic-stroke patients (N=77), age and education matched controls (N=22), and younger controls (N=20). We also included high resolution DWI data from the Human Connectome Project (HCP, N=37). Disconnection maps were estimated in eight ways: in-vivo stroke patients against 1) matched controls; 2) young controls; 3) HCP controls; and pseudo-lesion inferred disconnections (i.e. insert lesion into healthy data and remove intersecting connections) in 4) matched controls; 5) young controls; and 6) HCP controls. We also used two popular methods to obtain disconnection maps using the: 7) Brain Connectivity and Behaviour toolkit (BCB; Foulon et al., 2018) and 8) Network Modification tool (NeMo; Kuceyeski et al., 2013). Methods 2-8 were compared with Method 1 (the target) in two ways. First, we used correlation and cosine similarity of the disconnection values (summarised using AAL atlas). Second, we used permutation testing to determine if each Method differed to the target for each subject (quantifying the proportion of dissimilar subjects per method). The first analysis showed that every method was significantly different to the target, with the young controls being the closest. The second analyses showed that the proportion of dissimilar subjects were confined to the right hemisphere (6.5-54.5%), whereas the left hemisphere fared better (0-6.5%). Finally, we found no evidence of improved prediction performance using any disconnection method after accounting for lesion location. In summary, the results showed that inferred disconnection methods do not adequately reflect in-vivo disconnections; however, disconnection measures (from any method) did not account for unique variance in prediction analyses.