Data augmentation methods for machine-learning-based classification of bio-signals

Data augmentation methods for bio-signal classification are proposed. These methods improve recognition performance of human mental states showing intrinsic motivation from brain wave. Conventionally, data augmentation is used to image recognition research. Scaling, rotation, and distortion are applied to the original images to increase examples for machine learning. However, these augmentation methods are not effective for use with biological signals, as they involve spatial manipulation designed to represent the fluctuations of natural images. In the present study, we proposed four novel methods for data augmentation of biological signals. These methods are designed to represent variations inherent to bio-signals, especially for event-related signals. Electroencephalogram (EEG) data from participants engaged in an intrinsic motivation task were utilized to evaluate the feasibility of the proposed data augmentation methods. Our findings demonstrated that the proposed methods are particularly effective for improving prediction accuracy in small datasets.