A Single-Trial P300 Detector Based on Symbolized EEG and Autoencoded-(1D)CNN to Improve ITR Performance in BCIs

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Brain-computer interfaces (BCI) enable communication and control through neural signals.
• P300 event-related potentials are crucial for BCI applications like spellers.
• Existing P300 detection methods face challenges in speed, accuracy, and portability.

Purpose of the Study:

• To develop a high-performance, portable single-trial P300 detector for BCIs.
• To maximize the information transfer rate (ITR) while maintaining high recognition accuracy.
• To validate the algorithm's implementability on embedded systems.

Main Methods:

• A novel pre-processing stage involving EEG signal symbolization.
• An autoencoder model to emphasize temporal features.
• A seven-layer convolutional neural network (CNN) for classification.
• Acquisition from six EEG channels with low-complexity pre-processing (baseline correction, windsorizing, symbolization).

Main Results:

• Achieved an average ITR of 16.83 bits/min on a P300 speller dataset, outperforming state-of-the-art by 5.75 bits/min.
• Obtained an F1-Score of 51.78 ± 6.24% on the P300 speller dataset.
• Demonstrated an ITR of ~33 bit/min and 70.00% F1-Score in a prototype car driving experiment.
• Validated on an STM32L4 microcontroller, requiring <3.5 ms for classification with low resource usage (5.57% ROM, ~3% RAM).

Conclusions:

• The proposed single-trial P300 detector offers a significant advancement in BCI technology.
• The system achieves superior ITR and accuracy, suitable for real-time applications.
• The algorithm's efficiency and portability enable practical BCI implementation on embedded platforms.