A simple deep transfer learning model with feature alignment block for motor imagery decoding.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Motor imagery electroencephalography (MI-EEG) based brain-computer interfaces (BCIs) face challenges due to data scarcity and distribution shifts.
• Existing models often struggle with complex feature extraction and online alignment in MI-EEG decoding.

Purpose of the Study:

• To propose a novel 1-dimensional convolution-based deep transfer learning model with an embedded Feature Alignment block (1DC-DTL-FA).
• To address data scarcity and distribution shifts in MI-EEG decoding through an effective and simple architecture.

Main Methods:

• Developed a 1-dimensional convolution-based deep transfer learning model (1DC-DTL-FA) integrating multi-stage feature extraction, classification, and a Feature Alignment (FA) block.
• Utilized Neural Architecture Search (NAS) to automatically determine the optimal FA block position.
• Evaluated the model on the BCI 2000 and BCI IV2a datasets.

Main Results:

• The 1DC-DTL-FA model achieved superior accuracies of 89.80% on the BCI 2000 dataset and 82.96% on the BCI IV2a dataset.
• Demonstrated effective handling of complex feature extraction and online alignment.
• Outperformed existing state-of-the-art models in MI-EEG decoding.

Conclusions:

• The proposed 1DC-DTL-FA model offers a simple yet effective solution for MI-EEG decoding.
• This architecture successfully addresses data scarcity and distribution shifts, improving BCI performance.