[Paper Review] Transfer Learning for Motor Imagery Based Brain-Computer Interfaces: A Complete Pipeline

Dongrui Wu, Xue Jiang+4 more|arXiv (Cornell University)|Jul 3, 2020

EEG and Brain-Computer Interfaces51 references4 citations

TL;DR

This paper proposes a complete transfer learning (TL) pipeline for motor imagery-based brain-computer interfaces (BCIs), integrating TL across all key components—data alignment, spatial filtering, feature engineering, and classification—demonstrating that combining data alignment with multi-component TL significantly improves classification accuracy and reduces calibration effort in both cross-subject and cross-session settings.

ABSTRACT

Transfer learning (TL) has been widely used in motor imagery (MI) based brain-computer interfaces (BCIs) to reduce the calibration effort for a new subject, and demonstrated promising performance. While a closed-loop MI-based BCI system, after electroencephalogram (EEG) signal acquisition and temporal filtering, includes spatial filtering, feature engineering, and classification blocks before sending out the control signal to an external device, previous approaches only considered TL in one or two such components. This paper proposes that TL could be considered in all three components (spatial filtering, feature engineering, and classification) of MI-based BCIs. Furthermore, it is also very important to specifically add a data alignment component before spatial filtering to make the data from different subjects more consistent, and hence to facilitate subsequential TL. Offline calibration experiments on two MI datasets verified our proposal. Especially, integrating data alignment and sophisticated TL approaches can significantly improve the classification performance, and hence greatly reduces the calibration effort.

Motivation & Objective

• To address the high calibration burden in motor imagery-based BCIs due to individual EEG variability and non-stationarity.
• To overcome the limitation of prior TL approaches that only applied transfer learning to one or two components of the BCI pipeline.
• To propose and validate a comprehensive TL framework that integrates transfer learning across all major processing stages: data alignment, spatial filtering, feature engineering, and classification.
• To investigate the impact of explicit data alignment as a preprocessing step before TL to improve domain consistency across subjects.
• To evaluate the performance of multi-component TL in both cross-subject and cross-session BCI scenarios.

Proposed method

• Introduces a novel data alignment step using Euclidean alignment (EA) to reduce distribution differences between source and target domain EEG trials before spatial filtering.
• Applies transfer learning in spatial filtering using methods like CCSP and RCSP with TL-enhanced spatial filters.
• Integrates TL into feature engineering by leveraging domain-adaptive feature selection and extraction techniques.
• Employs TL in classification using algorithms like LDA, CLDA, and wAR, with adaptation regularization to improve generalization on limited labeled target data.
• Designs a complete pipeline (Figure 2) where TL is systematically applied across all stages: temporal filtering, data alignment, spatial filtering, feature engineering, and classification.
• Uses offline cross-subject and cross-session experiments on two MI EEG datasets to validate the pipeline.

Experimental results

Research questions

• RQ1 Can transfer learning be effectively applied across all major components of a motor imagery BCI pipeline, including data alignment, spatial filtering, feature engineering, and classification?
• RQ2 How does explicitly adding a data alignment step before spatial filtering affect the performance of subsequent transfer learning in BCIs?
• RQ3 Does integrating transfer learning across multiple components yield better classification performance than applying TL to only one or two components?
• RQ4 How does the proposed pipeline compare to conventional methods like CSP-LDA in terms of accuracy and calibration efficiency?
• RQ5 To what extent does the proposed TL pipeline reduce calibration effort in cross-subject and cross-session BCI applications?

Key findings

• The proposed complete TL pipeline, especially when combining data alignment with TL in spatial filtering and classification, achieved the highest average classification accuracy of 75.56% ± 0.05 on Dataset 2a in cross-session evaluation.
• The EA-CCSP-wAR method, which integrates data alignment and TL in spatial filtering and classification, outperformed all other approaches, showing a 21.29% improvement over the baseline CSP-LDA.
• Data alignment significantly enhanced TL performance: all methods with data alignment outperformed their counterparts without it, confirming its critical role in improving domain consistency.
• The integration of TL across multiple components was complementary, with multi-component TL consistently outperforming single-component TL across both datasets and evaluation scenarios.
• Even in non-stationary subjects, the proposed pipeline maintained robust performance, indicating its effectiveness in handling EEG signal variability across sessions.
• The results demonstrate that TL is highly effective not only in cross-subject but also in cross-session BCI applications, significantly reducing the need for lengthy calibration.