Poster E70, Thursday, August 22, 2019, 3:45 – 5:30 pm, Restaurant Hall
A shallow neural network to predict naming recovery in chronic stroke.
Barbara Khalibinzwa Marebwa1, Julius Fridriksson2, Chris Rorden3, Leonardo Bonilha1;1Department of Neurology, Medical University of South Carolina, 2Department of Communication Sciences and Disorders, University of South Carolina, 3Department of Psychology, University of South Carolina
Language impairment after stroke persists in up to 40% of chronic stroke survivors. Several studies have demonstrated that anodal tDCS coupled with speech therapy can lead to significant language improvement, improving the quality of life for stroke survivors. While the mechanism of action explaining this observation is yet to established, we aimed to test whether tDCS treatment and a combination of neuroanatomical and individual factors before treatment could predict maintenance of language 6 months after treatment. Seventy-one chronic stroke individuals with aphasia (40 women, mean age 59 ± 10.5 years) were randomized into two groups and underwent 15 picture-word matching language therapy sessions over a period of 3 weeks. One group (36 participants) received 1 mA anodal-tDCS while the other (35 participants) received sham-tDCS to the intact left temporoparietal region for the first 20 min of each session. The primary outcome was object naming improvement. Using post-processing methods of diffusion tensor imaging optimized for lesioned brains, we reconstructed individual structural whole-brain connectomes, and quantified the topological organization of each network using Newman’s modularity algorithm. We then trained a shallow pattern recognition neural net to dichotomize participants into two groups: those who performed better at 6 months compared to baseline vs those who performed worse, and a fitting neural network to determine a non-linear combination of behavioral and neuroanatomical factors that best predicted naming scores 6 months after therapy. In accordance with previous literature, we included the organization of the domain general and language specific left and right hemisphere networks, lesion load, the number of identified short-, mid- and long-range fiber connections as neuroanatomical factors, and age, time since stroke, and education as individual factors. We employed a 70:15:15 cross-validation scheme, meaning the training set contained 70% of the participants, the validation group, 15%, and testing was done on an independent sample- the remaining 15%. A grid search determined that 2 neurons and an 8-point threshold (6months - baseline) produced the highest classification accuracy (93% higher than a random model). We were able to classify the subjects into two groups: those who maintained naming performance (at least an 8-point improvement from baseline) vs. those who performed worse 6 months after treatment with up to 70%accuracy. For participants who performed better than the threshold (>8 points), right hemisphere modularity (93%), left hemisphere language network modularity (83%) and education (74%) were significant predictors associated with classification accuracy. For participants who performed worse than the threshold (<8 points), age (94%) and tDCS (81%) were significant predictors associated with classification accuracy. We were further able to predict PNT scores with 92% accuracy 6 months after treatment. We present simple neural networks that can not only determine participant language recovery up to 6-months after treatment, but also offers useful insight into disentangling factors necessary for language recovery, for instance the topological integrity of residual white matter networks, or factors leading to deteriorating language abilities for instance age, and the absence of an intervention e.g. tDCS.
Themes: Computational Approaches, Language Therapy
Method: White Matter Imaging (dMRI, DSI, DKI)