Data-driven Approaches to Improved Understanding of Brain-behaviour Relationships
Author | : Daniel Kristanto |
Publisher | : |
Total Pages | : 167 |
Release | : 2021 |
ISBN-10 | : OCLC:1286306158 |
ISBN-13 | : |
Rating | : 4/5 (58 Downloads) |
Book excerpt: The question of how the brain works has been the core interest of neuroscientists. To address this question, studies have been done to investigate the brain mechanism to receive, process, and transmit information. One of the approaches is to develop neuronal models to directly capture the microscopic activities and dynamics of the brain. Another approach emerges by explaining the brain organization to control human cognitive abilities as the mental expression of human behaviors which differ across individuals. This research direction has attracted scientists from different fields and is commonly known as brain-behavior relationships. In its development, this topic receives support from the advanced technology of brain imaging and behavioral tests, resulting in the availability of big and freely accessible data. Coupled with the development of mathematical techniques, recent perspective of brain-behavior study has shifted into data-driven approach which emphasizes on data exploration rather than initial hypothesis validation. Thus, some questions arise: (i) to ask whether data-driven approach in brain-behavior study, aside from providing new information, is also able to consistently capture the previously established knowledge at the level of single and multiple behaviors, ( ) towards its complexity, to inquire whether data-driven approach based on nonlinear model performs better than linear model in prediction and feature extraction, and ( i) to develop a general framework of fully nonlinear data-driven approach to improved the understanding of brain-behavior relationships. This thesis answers the first inquiry in Chapter 3 and Chapter 4 for single and multiple behaviors, respectively. In single behavior, a study on reading ability is performed by predicting the reading performance using brain structural and functional properties coming from brain area and connections measures. Aside from the prediction performances of reading ability, the study also interprets the features extracted from i linear data-driven method. Interestingly, the features presented as reading-related brain areas confirm the reading areas established in the literature and provide more understanding on the involvement of motor cortex in reading ability. Chapter 4 focuses on the comparison between cognitive abilities structure explored from the brain properties with the one established from hypothesis of performance covariance in multiple behaviors. Similarly, this study also shows the consistency of the data-driven approaches compared to hypothesis-driven approaches and reveals some interpretable dissimilarities in terms of cognitive ability structure. Research in Chapter 5 and Chapter 6 are carried out to address inquiries 2 and 3 focusing on the methods in data-driven approach. Chapter 5 systematically compares linear and nonlinear models to study the brain and behavior relationships. Engaging linear correlation and regression for the linear model and artificial neural network for nonlinear model, this study shows that the mapping between brain and behavior is better captured by nonlinear relationships. Further, the analyses on the features of both models reveal that the consistently captured features by both models are interpretable and relevant with the previous studies. In addition, the nonlinear model is also able to capture more interpretable features not identified in the linear model. Motivated by this finding, Chapter 6 proposes a framework to implement fully nonlinear data-driven approach towards improved understanding of brain-behavior relationships. This study recruits Graph Neural Networks, t-distributed Stochastic Neighbor Embedding, and kMeans Clustering to find brain clusters based on its structural properties. The proposed framework shows promising results. First, the structural clusters are partially consistent with the Resting State Networks identified from functional brain properties. Second, the structural properties of the brain belonging to each cluster similarly performs well to predict the performance of behavioral measures. Lastly, the wide variety of the iv behavioral measures which can be predicted by the clusters suggests that the clusters may reflect how the cognitive functions are organized in the brain. Altogether, this thesis shows the advantages of using data-driven approach to study brain-behavior relationships in single and multiple behavior levels. It also suggests that exploration using nonlinear approaches may provide better mapping and more interpretable features in brain and behavior studies.