What Causes Performance Degradation in Cross-Subject EEG Classification?

Authors:

Yihe Wang*(ywang145@charlotte.edu), Taida Li*(tli14@charlotte.edu), Yujun Yan, Wenzhan Song, Xiang Zhang(xiang.zhang@charlotte.edu)

Paper Link: [BMI@SMC 2026], Preprint

Takeaway Message

For performance degradation of cross-subject EEG classification, multi-class-per-subject (MCPS) EEG tasks are mainly affected by inter-subject variability, whereas single-class-per-subject (SCPS) EEG tasks are largely influenced by shortcut learning based on subject-specific features.

Two types of EEG task

Fig. 1: In MCPS tasks, labels are assigned to samples and may vary across samples within a subject (motor imagery, emotion recognition, sleep-stage classification, seizure detection). In SCPS tasks, each subject is assigned a single label, and all samples from that subject share the same label (Alzheimer’s disease, Parkinson’s disease, depression detection).

Requirements

The recommended requirements are specified as follows:

• Python==3.10
• einops==0.4.0
• matplotlib==3.7.0
• numpy==1.23.5
• pandas==1.5.3
• patool==1.12
• reformer-pytorch==1.4.4
• scikit-learn==1.2.2
• scipy==1.10.1
• sympy==1.11.1
• torch==2.5.1+cu121
• tqdm==4.64.1
• natsort~=8.4.0
• mne==1.9.0
• mne-icalabel==0.7.0
• h5py==3.13.0
• pyedflib==0.1.40
• linear_attention_transformer==0.19.1
• timm~=0.6.13
• transformers~=4.57.1

The dependencies can be installed by:

pip install -r requirements.txt

Data Preparation

The datasets used in this paper are all public datasets, and the data processing scripts are provided in data_processing/ folder, as well as readme files for each dataset.

Run Experiments

Before running, make sure you have all the processed datasets put under dataset/. Run meta_run.sh to reproduce all the experiments in paper. The results can be found in results/method_name/, and the checkpoints can be found in checkpoints/method_name/.