GTransU-CAP: Automatic labeling for cyclic alternating patterns in sleep EEG using gated transformer-based U-Net framework

Cyclic alternating pattern (CAP) sequences are composed of cycles of alternate activation phases (A-phases) and background phases. CAP A-phases can be further divided into three subtypes, which act as important bio-markers of sleep instability and are also associated with identifiable sleep pathologies. Thus, its accurate detection and identification is of great clinical interest and significance. To release the burden of sleep experts who manually perform this labeling task, several automatic detectors have been proposed, yet the characteristics of CAP have not been fully exploited to achieve a satisfactory performance. In this paper, we propose an automated method to detect A-phases and their subtypes using Transformer-based U-Net framework. In light of the long-span duration of A-phases, our method has intrinsic advantages as U-Net extracts local information while Transformer module provides global dependencies. We also use a curriculum-learning based training strategy to further improve the performance. The method is validated on the publicly available CAP Sleep Database. It obtains average F1 scores of 67.78% and 72.16% on 16 healthy subjects and 30 patients with nocturnal frontal lobe epilepsy respectively for A-phase detection, and the average macro F1-score is 59.5% for multi-class subtype classification. Compared with state-of-the-art methods, the proposed method achieves superior performance in these two CAP labeling tasks.

Automated summary

This paper proposes an automated method to detect CAP using a Transformer-based U-Net framework, achieving improved performance by extracting local information and providing global dependencies.