Presented By: Biomedical Engineering
BME Master's Thesis Defense - Ann Gu
Multi-Label Classification of Motor Tasks using fMRI Data
Within the past decade, predicting brain states by applying classification-based multi-voxel pattern analysis (MVPA) to functional magnetic resonance imaging (fMRI) data has become popular. In traditional classification-based MVPA, each class or label is modeled as having a unique spatial brain activity. Multi-Label classification is an emerging machine learning paradigm that can detect multiple classes, that are not necessarily mutually exclusive, in a single instance.
For this study, we extend a support vector machine (SVM) algorithm, a popular MVPA approach, to a multi-label algorithm that can detect both left and right hand tapping tasks simultaneously. Participants performed four tasks in a blocked experiment design: rest, right hand tapping, left hand tapping, and both hands tapping. We compare two training models with our multi-label data. One considers both hands tapping as a new class. The other considers both hands tapping as a positive instance of right and left hand tapping. Furthermore, we investigate the effects of SVM parameters on our algorithm’s performance. Our results demonstrates the feasibility of using a multi-label paradigm for motor task fMRI data. We discuss the capabilities and limitations of our approach and the potential to generalize to other fMRI task-based applications.
Chair: Scott Peltier
Within the past decade, predicting brain states by applying classification-based multi-voxel pattern analysis (MVPA) to functional magnetic resonance imaging (fMRI) data has become popular. In traditional classification-based MVPA, each class or label is modeled as having a unique spatial brain activity. Multi-Label classification is an emerging machine learning paradigm that can detect multiple classes, that are not necessarily mutually exclusive, in a single instance.
For this study, we extend a support vector machine (SVM) algorithm, a popular MVPA approach, to a multi-label algorithm that can detect both left and right hand tapping tasks simultaneously. Participants performed four tasks in a blocked experiment design: rest, right hand tapping, left hand tapping, and both hands tapping. We compare two training models with our multi-label data. One considers both hands tapping as a new class. The other considers both hands tapping as a positive instance of right and left hand tapping. Furthermore, we investigate the effects of SVM parameters on our algorithm’s performance. Our results demonstrates the feasibility of using a multi-label paradigm for motor task fMRI data. We discuss the capabilities and limitations of our approach and the potential to generalize to other fMRI task-based applications.
Chair: Scott Peltier
Explore Similar Events
-
Loading Similar Events...