STDA-Net: Spectrogram-Based Domain Adaptation for Sleep Stage Classification
The newly developed deep learning framework, STDA-Net, tackles the issue of classifying sleep stages across different datasets by utilizing two-dimensional spectrogram inputs within an unsupervised domain adaptation framework. This innovative method integrates a convolutional neural network for extracting features from spectrograms, a bidirectional long short-term memory module for modeling temporal aspects, and a domain-adversarial neural network to synchronize features from source and target datasets without the need for labeled target data during the training phase. The goal is to address the inconsistencies in EEG channel montages, sampling rates, recording settings, and participant groups that complicate accurate sleep staging across various datasets. Experiments are carried out to assess the effectiveness of this framework.
Key facts
- STDA-Net stands for Spectrogram-based Temporal Domain Adaptation Network.
- The framework uses a CNN for spectrogram-based feature extraction.
- A BiLSTM module models temporal dynamics of sleep.
- A DANN aligns source-to-target features without labeled target data.
- The method targets cross-dataset sleep stage classification.
- It addresses variability in EEG montages, sampling rates, environments, and populations.
- Most existing methods rely on one-dimensional EEG signals.
- The approach is published on arXiv with ID 2605.06736.
Entities
Institutions
- arXiv