SplitSEE: A Splittable Self-supervised Framework for Single-Channel EEG Representation Learning

TL;DR

This paper proposes two domain-specific modules that independently learn domain-specific representation and address the temporal-frequency tradeoff issue in conventional spectrogram-based methods and introduces a novel clustering loss to measure the information similarity.

Abstract

While end-to-end multi-channel electroencephalography (EEG) learning approaches have shown significant promise, their applicability is often constrained in neurological diagnostics, such as intracranial EEG resources. When provided with a single-channel EEG, how can we learn representations that are robust to multi-channels and scalable across varied tasks, such as seizure prediction? In this paper, we present SplitSEE, a structurally splittable framework designed for effective temporal-frequency representation learning in single-channel EEG. The key concept of SplitSEE is a self-supervised framework incorporating a deep clustering task. Given an EEG, we argue that the time and frequency domains are two distinct perspectives, and hence, learned representations should share the same cluster assignment. To this end, we first propose two domain-specific modules that independently learn domain-specific representation and address the temporal-frequency tradeoff issue in conventional spectrogram-based methods. Then, we introduce a novel clustering loss to measure the information similarity. This encourages representations from both domains to coherently describe the same input by assigning them a consistent cluster. SplitSEE leverages a pre-training-to-fine-tuning framework within a splittable architecture and has following properties: (a) Effectiveness: it learns representations solely from single-channel EEG but has even outperformed multi-channel baselines. (b) Robustness: it shows the capacity to adapt across different channels with low performance variance. Superior performance is also achieved with our collected clinical dataset. (c) Scalability: With just one fine-tuning epoch, SplitSEE achieves high and stable performance using partial model layers.

Figures (5)

• Figure 1: The scatter plot presents the seizure prediction results of SplitSEE compared with single-channel baselines on two public EEG datasets. Points located in the upper-left quadrant indicate that our single-channel method maintains robust performance across various EEG inputs with lower variance and higher average accuracy. These results underscore the superior performance and reliability of SplitSEE in seizure prediction.
• Figure 2: SplitSEE overview. The architecture integrates a two-phase learning approach: an initial self-supervised pre-training phase followed by a fine-tuning phase. It comprises three representation learning modules: two domain-specific feature learning modules for temporal-frequency independent learning, and a cross-domain feature clustering module. Moreover, the splittable architecture allows seamless adaptation to various tasks using selected layers. Each domain-specific module is trained by InfoNCE loss function, and the deep clustering for alignment is performed by maximizing mutual information (MI).
• Figure 3: Description of (a) seizure prediction and (b) sleep stage scoring tasks. (a) Seizure prediction task involves predicting epileptic seizures using EEG data designed to alert patients before an event. (b)Sleep stage scoring task categorizes sleep patterns into distinct stages, facilitating an understanding of sleep quality and disorders.
• Figure 4: UMAP visualization of raw data and features in (a) CHB-MIT and (b) SleepEDF datasets. The strength of the proposed method lies in the clear clustering boundaries observed in both training and test data for both tasks.
• Figure 5: Results of self-supervised training loss (left) and fine-tuned classification accuracy (right) on the (a) HUH and (b) SleepEDF datasets. Accuracy immediately rosed to around 80% after the first fine-tuning epoch.