Multimodal Network With Accuracy-Based Inverse Loss Weighting and Multimodal Ensemble Classification for EEG-Based Sleep Stage Classification

Koohong Jung, Moogyeong Kim, Wonzoo Chung

Abstract: In this paper, a multimodal network with a novel loss weighting scheme, accuracy-based inverse loss weighting, is proposed for sleep stage classification based on electroencephalography (EEG). Although multimodal sleep stage classification approaches that extract features from time and time-frequency domains have shown promising results, they overlook underfitted modality, which can cause poor generalization. To encourage learning from underfitted modality, we propose to assign weight to each domain loss, which is inversely proportional to the validation accuracy of the domain. To effectively capture the diverse temporal dynamics of sleep transitions, a neural network architecture, which employs bidirectional temporal convolution, is proposed. In addition, a novel ensemble method that aggregates the ensembles of all multimodal features is proposed to improve the overall classification performance rather than relying on a single-domain output ensemble as in existing methods. The performance of the proposed method is evaluated using the SleepEDF-20 and SleepEDF-78 datasets. Experimental results show that the proposed method exhibits superior classification performance compared with existing algorithms.

External IDs:doi:10.1109/access.2025.3647958