Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Exploring the Relationship between Brain Hemisphere States and Frequency Bands through Classical Machine Learning and Deep Learning Optimization Techniques with Neurofeedback

Robiul Islam, Dmitry I. Ignatov, Karl Kaberg, Roman Nabatchikov

Abstract

This study investigates the performance of classifiers across EEG frequency bands, evaluating efficient class prediction for the left and right hemispheres using various optimisers. Three neural network architectures a deep dense network, a shallow three-layer network, and a convolutional neural network (CNN) are implemented and compared using the TensorFlow and PyTorch frameworks. Adagrad and RMSprop optimisers consistently outperformed others across frequency bands, with Adagrad excelling in the beta band and RMSprop achieving superior performance in the gamma band. Classical machine learning methods (Linear SVM and Random Forest) achieved perfect classification with 50--100 times faster training times than deep learning models. However, in neurofeedback simulations with real-time performance requirements, the deep neural network demonstrated superior feedback-signal generation (a 44.7% regulation rate versus 0% for classical methods). SHAP analysis reveals the nuanced contributions of EEG frequency bands to model decisions. Overall, the study highlights the importance of selecting a model dependent on the task: classical methods for efficient offline classification and deep learning for adaptive, real-time neurofeedback applications.

Exploring the Relationship between Brain Hemisphere States and Frequency Bands through Classical Machine Learning and Deep Learning Optimization Techniques with Neurofeedback

Abstract

This study investigates the performance of classifiers across EEG frequency bands, evaluating efficient class prediction for the left and right hemispheres using various optimisers. Three neural network architectures a deep dense network, a shallow three-layer network, and a convolutional neural network (CNN) are implemented and compared using the TensorFlow and PyTorch frameworks. Adagrad and RMSprop optimisers consistently outperformed others across frequency bands, with Adagrad excelling in the beta band and RMSprop achieving superior performance in the gamma band. Classical machine learning methods (Linear SVM and Random Forest) achieved perfect classification with 50--100 times faster training times than deep learning models. However, in neurofeedback simulations with real-time performance requirements, the deep neural network demonstrated superior feedback-signal generation (a 44.7% regulation rate versus 0% for classical methods). SHAP analysis reveals the nuanced contributions of EEG frequency bands to model decisions. Overall, the study highlights the importance of selecting a model dependent on the task: classical methods for efficient offline classification and deep learning for adaptive, real-time neurofeedback applications.

Paper Structure

This paper contains 74 sections, 9 equations, 12 figures, 34 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Experimental Design
  3. Machine Learning Methods
  4. Evaluation Metrics
  5. Hypothesis Testing
  6. Verification
  7. Visual Representation and Interpretable Machine Learning
  8. Method
  9. Classical Machine Learning Methods
  10. Related work
  11. EEG and language comprehension
  12. Visual Stimuli
  13. Left and right hemispheres
  14. Batch Normalization in deep networks
  15. Neurofeedback
  16. ...and 59 more sections

Figures (12)

  • Figure 1: A line drawing of a transparent cube (known as the Necker cube), with opposite sides drawn parallel, so that the perspective is ambiguous
  • Figure 2: Leonardo da Vinci, Portrait of Lisa Gherardini (known as the Mona Lisa), c. 1503--19, oil on poplar panel
  • Figure 3: Overview of EEG Data Processing Pipeline
  • Figure 4: SHAP plots for the $\beta$ frequency using the RMSProp optimizer, with the Necker cube dataset as input. Feature $F$ corresponds to the time $F/250$ s from the beginning of image presentation.
  • Figure 5: Accuracy and loss plot for $\beta$ rhythm using the RMSProp optimizer, with the Necker cube dataset as input
  • ...and 7 more figures