SSL-SE-EEG: A Framework for Robust Learning from Unlabeled EEG Data with Self-Supervised Learning and Squeeze-Excitation Networks

TL;DR

SSL-SE-EEG tackles the challenge of reducing labeled data needs in EEG analysis while improving robustness to noise. By transforming EEG into CNN-compatible 2D RGB images and integrating Squeeze-and-Excitation networks into a self-supervised learning framework with a two-stage training process, the approach learns transferable representations and achieves strong cross-dataset performance. Key contributions include the EEG-to-image representation, the SE-Net–augmented SSL pipeline, and evidence of high accuracy on multiple public EEG datasets alongside minimal power overhead, making it suitable for real-time, wearable BCIs. The work demonstrates significant potential for scalable, energy-efficient neural engineering and neural-interface applications in cognitive state monitoring and neurorehabilitation.

Abstract

Electroencephalography (EEG) plays a crucial role in brain-computer interfaces (BCIs) and neurological diagnostics, but its real-world deployment faces challenges due to noise artifacts, missing data, and high annotation costs. We introduce SSL-SE-EEG, a framework that integrates Self-Supervised Learning (SSL) with Squeeze-and-Excitation Networks (SE-Nets) to enhance feature extraction, improve noise robustness, and reduce reliance on labeled data. Unlike conventional EEG processing techniques, SSL-SE-EEG} transforms EEG signals into structured 2D image representations, suitable for deep learning. Experimental validation on MindBigData, TUH-AB, SEED-IV and BCI-IV datasets demonstrates state-of-the-art accuracy (91% in MindBigData, 85% in TUH-AB), making it well-suited for real-time BCI applications. By enabling low-power, scalable EEG processing, SSL-SE-EEG presents a promising solution for biomedical signal analysis, neural engineering, and next-generation BCIs.

Figures (6)

• Figure 1: (a) EEG workflow highlighting challenges in data labeling and processing. (b) Trade-off between model accuracy and labeling cost across ML paradigms, showing how SSL-SE-EEG achieves high performance with reduced labeled data.
• Figure 2: Overview of the proposed SSL-SE-EEG pipeline, which consists of two steps. Step 1 involves preprocessing EEG signals into 2D image representations. Step 2 integrates self-supervised learning with SE-Nets through a two-stage process: Stage 1 applies contrastive learning using a modified encoder (BE+SE) and a projection head (PH) to maximize feature diversity; Stage 2 fine-tunes the network on a smaller labeled dataset for classification, leveraging learned representations for robust inference.
• Figure 3: Illustration of EEG-to-2D image conversion: (a) Raw EEG signals are segmented into 2-second windows, where 50 ms temporal segments are flattened into 1D columns to construct a 2D image. (b) Examples from TUH-AB and BCI-IV datasets demonstrate the effectiveness of this transformation for CNN-based feature extraction and classification.
• Figure 4: Modified Base Encoder with SE-Net. SE-blocks enhance EEG feature sensitivity via global pooling and fully connected layers.
• Figure 5: Generalization performance of SSL-SE-EEG across datasets. The model is pretrained on MBD or TUH-AB and tested on unseen subsets, showing strong cross-dataset transfer with minimal accuracy drop.