Network Level Language Processing Deficits in Post-stroke Aphasia: Relating Machine Learning Based Classification and Prediction with Resting-state FMRI

Previous work has provided strong evidence that resting-state fMRI functional connectivity within the dual-stream speech network is a potential measurement of post-stroke language abilities and aphasia symptoms. However, the effects of post-stroke hyper- versus hypo-functional connectivity on aphasia symptoms is not well understood. Also, the gap between the laboratory studies of neuroimaging data predicting aphasia, and a practical model that can be deployed as a medical application remains a major obstacle in the field. To address the first question in dissertation, experiment 1 examines two small independent resting-state fMRI datasets to better characterize post-stroke functional connectivity in relation to the direction, magnitude, and distance of the connectivity differences between 28 chronic left-hemisphere stroke survivors with aphasia and 28 neurotypical matched control participants collected at another site. To overcome the second gap in the literature, experiment uses a large Human Connectome Project aging dataset of neurotypical control participants and the largest post-stroke aphasia fMRI database available to yield an accurate prediction of a patient’s current language abilities and aphasia diagnosis, thereby making a meaningful step towards filling the gap between laboratory level research and the real-world application of machine learning of neuroimaging data. For experiment 1, the results indicate the following: 1) long distance functional connectivities were more affected post-stroke than shorter connectivities and post-stroke hypo-functional connectivity more common than post-stroke hyper-functional connectivity; 2) intra-right hemisphere functional connectivity in the stroke group was not significantly higher or lower than the control group; 3) greater bilateral ventral stream functional connectivities were significant predictors of better language abilities in several tasks and lower overall aphasia severity, while greater inter-hemisphere connectivities of the right dorsal stream negatively affected language abilities. Experiment 2 concluded that a transfer learning strategy could generate practical models in both classification and regression tasks in post-stroke aphasia, even if using a relatively small dataset. Together, both experiments indicate the value and unique contributions of using network-level functional connectivity analyses to explore the underlying neural mechanisms in post-stroke aphasia, and provide evidence in support of the feasibility of developing clinically relevant, fast, and accurate aphasias evaluations via MRI data.