Related Experiment Video
Updated: May 28, 2025
08:51
Statistical Modelling of Cortical Connectivity Using Non-invasive Electroencephalograms
Published on: November 1, 2019
5.6K
Stochasticity as a solution for overfitting-A new model and comparative study on non-invasive EEG prospects
Yousef A Radwan1, Eslam Ahmed Mohamed1, Donia Metwalli1
1Center for Informatics Science (CIS), School of Information Technology and Computer Science, Nile University, Sheikh Zayed City, Giza, Egypt.
Frontiers in Human Neuroscience
|February 10, 2025
Summary
Inner speech Brain-Computer Interfaces (BCIs) show promise. A new classifier, BruteExtraTree, achieved 46.6% accuracy in subject-dependent scenarios, outperforming current methods.
Area of Science:
- Neuroscience
- Machine Learning
- Brain-Computer Interfaces
Background:
- Inner speech signals are crucial for practical Brain-Computer Interface (BCI) applications but are challenging to decipher.
- Subject-dependent variability and high noise levels hinder BCI development.
Purpose of the Study:
- To evaluate Machine Learning (ML) and Deep Learning (DL) models for inner speech BCI.
- To address overfitting and improve accuracy in inner speech BCI.
Main Methods:
- Utilized a publicly available dataset with popular preprocessing methods for feature extraction.
- Proposed and evaluated a novel classifier, BruteExtraTree, based on ExtraTreeClassifier.
- Compared BruteExtraTree against Deep Learning models like ShallowFBCSPNet.
Main Results:
- BruteExtraTree achieved 32% accuracy in subject-independent scenarios, matching the best DL model.
- BruteExtraTree surpassed state-of-the-art with 46.6% average per-subject accuracy in subject-dependent scenarios.
- Results indicate potential for a new paradigm in inner speech BCI, inspired by LLM pretraining.
Conclusions:
- Inner speech BCI shows significant potential, especially in subject-dependent configurations.
- Further advancements require improved data recording or noise reduction techniques for practical subject-independent applications.
