نهمین کنگره نقشه برداری مغز

نمایش مقاله

Title:

Hybrid Signal Processing and Machine Learning Approaches for Cross-Dataset EEG-Based BCI Analysis

Subject:

Brain–Computer Interfaces (BCIs)

Presenter Author:
Name:	Morteza Shahedifar

Author:
Name 1:	Morteza Shahedifar - Corresponding - Presenter
Affiliation 1:	Electrical and Computer Eng. University of Tabriz, Iran

Name 2:	Alireza Siadatan
Affiliation 2:	Associate Professor, Senior Member of IEEE, University of Toronto, Canada

Name 3:
Affiliation 3:

Name 4:
Affiliation 4:

Name 5:
Affiliation 5:

Name 6:
Affiliation 6:

Name 7:
Affiliation 7:

Name 8:
Affiliation 8:

Name 9:
Affiliation 9:

Name 10:
Affiliation 10:

Abstract

Original Research

introduction :

Electroencephalography (EEG)-based brain computer interfaces (BCIs) are commonly developed and evaluated using data from a single dataset acquired under controlled experimental conditions. While this practice simplifies model development, it does not adequately address a critical limitation of EEG-based BCIs: poor generalization across datasets collected using different recording protocols, hardware configurations, and subject populations. In practical scenarios, models trained on one dataset often experience a substantial performance drop when applied to unseen datasets, even when the underlying task remains the same. Despite its importance for real-world BCI applications, cross-dataset generalization has received limited systematic attention in the existing literature.

Methods:

In this study, a cross-dataset analysis framework is developed to investigate the generalization behavior of EEG-based BCI models using publicly available motor imagery datasets. EEG recordings from benchmark sources, including PhysioNet and BCI Competition datasets, are used without additional data collection. A hybrid processing pipeline is employed, combining conventional EEG signal processing with machine learning-based classification. Preprocessing steps include band-pass filtering and spatial feature extraction using common spatial patterns to reduce inter-dataset variability at the signal level. The extracted features are then used as input to compact learning-based models, including convolutional and recurrent neural architectures. Model training is performed on one dataset and evaluated on a different dataset to explicitly assess cross-dataset performance. Additionally, lightweight adaptation strategies based on feature normalization and limited fine-tuning are examined to mitigate distribution mismatch between datasets.

Results:

Cross-dataset evaluation results demonstrate that models trained and tested within a single dataset achieve notably higher performance compared to models evaluated across datasets. A significant reduction in classification accuracy is observed when dataset-specific models are directly transferred to unseen datasets. The proposed hybrid approach shows improved robustness under cross-dataset conditions, particularly when classical signal processing is used to constrain the feature space prior to learning-based modeling. The results indicate that incorporating domain-informed preprocessing can partially alleviate performance degradation and lead to more consistent class-wise behavior across datasets.

Conclusion:

The results of this study highlight cross-dataset generalization as a fundamental challenge in EEG-based BCI research. The results suggest that hybrid signal processing and machine learning approaches offer a practical and reproducible strategy for improving robustness when models are applied beyond the dataset used for training. By relying exclusively on public EEG datasets and standard simulation environments, the proposed framework provides a realistic basis for future methodological investigations in EEG-based brain mapping and BCI development.

Keywords:

BCI; Motor Imagery; Machine Learning; Signal Processing; Data Generalization