A hybrid autoencoder framework of dimensionality reduction for brain-computer interface decoding.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Brain-computer interfaces (BCIs) face challenges with high-dimensional neural data from large-scale recordings.
• Dimensionality reduction is crucial for efficient preprocessing of neural features like electrocorticography (ECoG) and local field potential (LFP).

Purpose of the Study:

• To propose a novel deep learning framework for effective dimensionality reduction of neural features.
• To enhance the performance and stability of BCIs by optimizing neural signal processing.

Main Methods:

• Implemented a high-performance autoencoder with convolutional layers for spatial and frequency dimensions.
• Utilized bottleneck long short-term memory (LSTM) layers for temporal dimension processing.
• Integrated a fully connected layer for regularization during autoencoder training.

Main Results:

• The proposed deep learning method significantly outperformed traditional techniques like KPCA, PLS, PSID, and LFADS.
• Reduced dimensionality features were successfully applied to various BCI decoders with no significant loss in decoding performance.

Conclusions:

• The novel deep learning framework offers a reliable tool for efficient neural signal dimensionality reduction.
• The method's performance and robustness are expected to improve decoding accuracy and long-term stability in online BCI systems.