A novel dual-stream time-frequency contrastive pretext tasks framework for sleep stage classification

Sergio Kazatzidis,Siamak Mehrkanoon

2023-12-15

Abstract:Self-supervised learning addresses the challenge encountered by many supervised methods, i.e. the requirement of large amounts of annotated data. This challenge is particularly pronounced in fields such as the electroencephalography (EEG) research domain. Self-supervised learning operates instead by utilizing pseudo-labels, which are generated by pretext tasks, to obtain a rich and meaningful data representation. In this study, we aim at introducing a dual-stream pretext task architecture that operates both in the time and frequency domains. In particular, we have examined the incorporation of the novel Frequency Similarity (FS) pretext task into two existing pretext tasks, Relative Positioning (RP) and Temporal Shuffling (TS). We assess the accuracy of these models using the Physionet Challenge 2018 (PC18) dataset in the context of the downstream task sleep stage classification. The inclusion of FS resulted in a notable improvement in downstream task accuracy, with a 1.28 percent improvement on RP and a 2.02 percent improvement on TS. Furthermore, when visualizing the learned embeddings using Uniform Manifold Approximation and Projection (UMAP), distinct clusters emerge, indicating that the learned representations carry meaningful information.

Signal Processing,Machine Learning

What problem does this paper attempt to address?

The paper primarily aims to address the following issues:

1. **Reducing the need for large amounts of labeled data**: In supervised learning methods, a large amount of labeled data is often required for training, and obtaining this data is both time-consuming and costly. This is especially true in the field of electrophysiology (EEG) research, where the data labeling process is very resource-intensive.

2. **Improving sleep stage classification tasks**: By enhancing the accuracy of sleep stage classification to assist in detecting sleep disorders such as sleep apnea. Traditional deep learning methods (such as convolutional neural networks) are effective but still rely heavily on large amounts of labeled data.

To address the above challenges, the authors propose a new framework based on self-supervised learning. Specifically, this framework includes the following key points:

- **Dual-stream time-frequency contrastive pre-training task architecture**: This is a pre-training task design that combines time-domain and frequency-domain information, aiming to capture the features of EEG signals from different perspectives.

- **Introduction of a novel pre-training task—Frequency Similarity (FS)**: The FS task is conducted in the frequency domain and aims to assess the spectral similarity between different windows. This task is combined with existing time-domain pre-training tasks (Relative Positioning RP and Temporal Shuffling TS) to enhance data representation capabilities.

- **Performance validation**: Through sleep stage classification experiments on the Physionet Challenge 2018 dataset, it is demonstrated that the proposed FS task can significantly improve the performance of downstream tasks, especially in scenarios with a small number of labeled samples.

In summary, this study aims to improve the effectiveness of self-supervised learning in electrophysiological data processing, particularly for sleep stage classification tasks, by introducing the frequency similarity pre-training task and designing a dual-stream architecture to integrate time-domain and frequency-domain information.