When planning neurosurgical interventions, identifying the brain areas involved in speech and language is of paramount importance. Some people are unable to participate in standard language localisation procedures due to intellectual disability or other neurodevelopmental reasons. This group of people is most at risk to have an abnormal distribution of the brain areas supporting speech and language. Due to the need to identify language lateralisation in these subjects, the thesis investigated whether it is possible to identify the individual's dominant hemisphere for speech from structural brain images and the functional connectivity of brain regions at rest. We used open access data from healthy individuals (N = 962) from the Human Connectome Project aged between 22 and 36 years. Each participant attended two functional magnetic resonance imaging scanning on different days, during which structural brain images, rasting state functional images and task based functional images were captured, including during performance of a language task involving short auditory stories followed by a forced-choice question with two alternatives. The resting-state functional connectivity matrix was analysed using graph theory measures. We averaged each structural measure within each brain parcel. The structural and network measures represented the machine learning input variables, which were used to teach the algorithm to identify the language lateralization, which we had previously determined for each participant based on brain activations during the language task. In this way, we obtained several predictive models of language lateralisation. We examined the reliability of the predictions on a separate sample. We complemented the results with an exploratory analysis in which we tried to increase the predictive power of the models by selecting the most relevant variables. Predictive models of atypical and right language lateralisation showed poor predictive power on the test sample as they showed low accuracy in identifying people with atypical and right language lateralisation. Models based on network measures and models based on the language network proved to be more promising. The predictive models developed are not suitable for clinical use due to their lack of predictive power. In particular, our findings should be considered as guidelines for further research.
|
