조건부 변분 오토인코더를 활용한 교차 피험자 기반 생성형 데이터 증강을 통한 뇌-컴퓨터 인터페이스 성능 향상
Inter-Subject Generative Data Augmentation Using Conditional Variational Autoencoder for Enhanced BCI Performance
1
조회
0
다운로드
0
인용
0
참고문헌
키워드
초록
Reliable EEG-based brain-computer interface (BCI) performance is often hindered by data scarcity and the non-stationarity of EEG signals. To address these challenges, this study proposes an inter-subject data augmentation approach using a conditional variational autoencoder (cVAE). The cVAE was trained on pooled data from 29 subjects and conditioned on both task and subject labels, enabling it to learn latent representations that disentangle task-relevant features shared across subjects from subject-specific characteristics. This structure allows the model to generate diverse and realistic synthetic EEG that reflects population-level patterns while preserving the unique traits of the target subject. Classification performance was evaluated by comparing a baseline model trained on original data with models trained using different augmentation strategies. The artificial data generation (ADG) method—a technique that rearranges temporal segments to create new signals—and the intra-subject cVAE approach did not yield meaningful performance gains, indicating limited improvement in data variability. In contrast, the proposed inter-subject cVAE increased average classification accuracy from 72.2±8.7% to 87.6±7.2% (15% gain). These results indicate that leveraging population information through generative modeling is an effective strategy for enhancing BCI robustness in data-limited environments
Abstract
Reliable EEG-based brain-computer interface (BCI) performance is often hindered by data scarcity and the non-stationarity of EEG signals. To address these challenges, this study proposes an inter-subject data augmentation approach using a conditional variational autoencoder (cVAE). The cVAE was trained on pooled data from 29 subjects and conditioned on both task and subject labels, enabling it to learn latent representations that disentangle task-relevant features shared across subjects from subject-specific characteristics. This structure allows the model to generate diverse and realistic synthetic EEG that reflects population-level patterns while preserving the unique traits of the target subject. Classification performance was evaluated by comparing a baseline model trained on original data with models trained using different augmentation strategies. The artificial data generation (ADG) method—a technique that rearranges temporal segments to create new signals—and the intra-subject cVAE approach did not yield meaningful performance gains, indicating limited improvement in data variability. In contrast, the proposed inter-subject cVAE increased average classification accuracy from 72.2±8.7% to 87.6±7.2% (15% gain). These results indicate that leveraging population information through generative modeling is an effective strategy for enhancing BCI robustness in data-limited environments