An Efficient Self-Supervised Framework for Long-Sequence EEG Modeling

• 2025-06-06: Repository anonymized to comply with the triple-blind reviewing policy of the ICDM conference.
• 2025-09-21: Paper accepted at the AI4TS Workshop @ ICDM 2025.
• Paper is in proceedings. Preprint available on arXiv.

Overview

EEGM2 consists of two stages:

1. Self-supervised pretraining
2. Downstream classification tasks

Pretrained models are provided in the Pretrained_EEGM2/ folder.

Dataset structure and preparation guidelines are described in the Datasets/ folder.

Model parameters and training settings can be configured in config.json (for pretraining) and config_downstream.json (for downstream tasks).

Setup

Instructions refer to Unix-based systems (e.g., Linux, macOS).

This code has been tested with Python 3.10.16. For users who prefer using Conda, environment YAML files are provided.

Create an environment from a specific file:

conda env create -f Env-requirement/Mamba2-env20250206.yml

Run Commands

1. Pretrain EEGM2 on TUAB (e.g., 30-second segments)

python SelfSupervised.py --config config.json

2. Run downstream classification on Crowdsourced dataset

python Downstream.py --config config_downstream.json

Results Summary (TUAB Dataset)

Balanced Accuracy and AUROC across varying input sequence lengths.

Duration	Model	Balanced ACC	AUROC
10 Seconds	EEGM2 (Light)	79.14 ± 0.21	0.8559 ± 0.000
	EEGM2 (Fine)	80.87 ± 0.54	0.8864 ± 0.000
30 Seconds	EEGM2 (Light)	78.97 ± 0.25	0.8575 ± 0.000
	EEGM2 (Fine)	81.71 ± 0.12	0.8932 ± 0.000
60 Seconds	EEGM2 (Light)	76.94 ± 0.33	0.8257 ± 0.000
	EEGM2 (Fine)	80.68 ± 0.45	0.8803 ± 0.000
100 Seconds	EEGM2 (Light)	74.57 ± 0.27	0.7986 ± 0.000
	EEGM2 (Fine)	81.08 ± 0.28	0.8869 ± 0.000