Brain computer interface based
Download as PPTX, PDF1 like828 views
1) The document presents research on using EEG signals to predict reaching targets during two experiments. 2) It describes the components of a BCI system and the challenges of using EEG data, which can be contaminated by artifacts from eye and muscle movements. 3) The researchers used independent component analysis and other techniques to remove artifacts from the EEG data before extracting features and using classification algorithms to decode reaching targets.
![• The first step of our artifact removal approach is to run
independent component analysis (ICA) on the EEG data. ICA is
a statistical technique which takes recordings from an array of
sensors and determines a set of source signals which are
maximally independent according to a specified measure of
statistical independence. ICA posits a data model X = AS,
where X are stacked row vectors of data recorded from
individual sensors, A is a matrix of mixing weights, and S are
the stacked row vectors of statistically independent source
vectors.
• There are a variety of ICA algorithms which primarily differ in
the independence measure used. For artifact removal, we use
InfoMax ICA [1], [9], which aims to minimize mutual
information between sources by maximizing entropy. For ICA
calculations and component visualization, use EEGLAB .
Independent Component Analysis(ICA)](https://image.slidesharecdn.com/braincomputerinterfacebased-101126080850-phpapp02/85/Brain-computer-interface-based-15-320.jpg)
Brain computer interface based
- 1. Brain Computer Interface Based on EEG Siganl Predicting Reaching Targets from Human EEG
- 2. Agenda • Introduction. • Problems To Build BCI System. • Goal. • BCI System. • Experiment One. • Data Processing. • Overview. • Experiment Two. • Conclusion.
- 3. Introduction • BCIs are being developed for a variety of applications ranging from assistive technologies for patients with motor disabilities to entertainment devices. • Extracting reach information from brain signals is of great interest to the fields of brain-computer interfaces (BCIs) and human motor control. To date, most work in this area has focused on invasive intracranial recordings; however, successful decoding of reach targets from noninvasive electroencephalogram (EEG) signals would be of great interest
- 4. • Show that EEG signals contain sufficient information to decode target location during a reach (Experiment 1) and during the planning period before a reach (Experiment 2). Introduction
- 5. • There are two main reasons why EEG data is rarely used for reaching tasks. The first drawback and perhaps most significant reason is due to contamination from artifacts.EEG electrodes sense all electrical activity at the scalp and therefore record not only electrical signals generated by brain activity but also electrical activity generated by eye movements and contraction of head and neck muscles. The second drawback of EEG is that it records the aggregate activity of large numbers of neurons. Thus, it is generally thought that EEG does not extract sufficiently detailed information to reconstruct complex movements Problems To Build BCI System
- 6. • The large number of degrees of freedom, high complexity, and speed of natural movement pose particular challenges to building BCI systems of this type. Problems To Build BCI System
- 7. • One goal of BCI research is to develop systems capable of decoding neural representations of natural movement planning and execution. • Our goal is to show that with appropriate artifact removal, signal processing, and machine learning human EEG carries sufficient information about reach intention to decode reaching targets from the EEG signal during movement planning and execution. Goal
- 8. • All BCI systems share the same set of underlying components, which can be broken down into three main segments: brain signal acquisition, brain state decoding, and computer-mediated performance of a task. • Some BCIs decode the brain state into a set of discrete classes such as yes/no commands, while other BCIs decode continuous data such as a reaching trajectory. BCI System
- 9. • EEG and three-dimensional (3-D) movement data were simultaneously recorded during a cued reaching task in a darkened room. • Subjects (four male, average age 44, all right- handed) sat in a comfortable chair and used their right index finger to reach for one of three target LEDs situated to the left (L), center (C), and right (R) of the subject’s sternum. • An individual trial began with LED illumination. Subjects were instructed to make an immediate and natural reach to the target. Experiment One
- 10. • The LED stayed on until the target was touched, or for 2,500 ms in case of target miss. After touching the target, subjects were instructed to pause briefly and then return to the starting position. Experiment One
- 11. • Reduce The Dimensionality. • EEG Artifact Detection And Removal. • Feature Extraction. • Classification Approach. Data Preprocessing
- 12. • Reduce the dimensionality by cutting down the number of channels from 256 to 140 or 150. Reduce The Dimensionality
- 13. • EEG data recorded during eye or other muscle movement typically contain movement artifacts. • Eye blinks and other movements generate a varying electrical field that propagates throughout the head and is picked up by scalp electrodes. Muscles generate electrical activity when they contract. • Muscle tension in scalp, face, and neck muscles generates signals which also propagate throughout the scalp adding electromyographic (EMG) activity to electrodes near the muscle insertion. EEG Artifact Movements
- 14. • The standard approach to handling artifacts in EEG research is to avoid them by developing experiments which restrict movement as much as possible, and then discarding trials which contain movement artifacts. This approach, however, is impractical for natural reaching tasks, in which eye and muscle movement are generally unavoidable. • Because reach data is especially susceptible to contamination from EEG artifacts that could potentially lead to erroneous classification rates, we employ an especially conservative artifact removal procedure, which is described in the following slide. EEG Artifact Detection And Removal
- 15. • The first step of our artifact removal approach is to run independent component analysis (ICA) on the EEG data. ICA is a statistical technique which takes recordings from an array of sensors and determines a set of source signals which are maximally independent according to a specified measure of statistical independence. ICA posits a data model X = AS, where X are stacked row vectors of data recorded from individual sensors, A is a matrix of mixing weights, and S are the stacked row vectors of statistically independent source vectors. • There are a variety of ICA algorithms which primarily differ in the independence measure used. For artifact removal, we use InfoMax ICA [1], [9], which aims to minimize mutual information between sources by maximizing entropy. For ICA calculations and component visualization, use EEGLAB . Independent Component Analysis(ICA)
- 16. • From the cleaned EEG data we extract a set of eight different feature vectors which were useful for classification. Feature 1: Autoregressive Modeling. Feature 2: Power Estimates Using A Filter Bank. Feature 3: Cleaned EGG Data. Feature 4: Discrete Wavelet Transform. Features 5–8: ICA Transforms. Feature Extraction
- 17. • Our classification approach involves creating individual classifiers based on each of the eight different feature vectors, and then combining these individual classifiers into a meta-classifier. Classification Approach
- 18. • Multinomial logistic classification. • Combining classifiers. Classification Approach
- 20. • Reaching target error rates were somewhat high for the plan segment (ranging from 34% to 50% error) and generally lower for the move segment (ranging from 26% to 46% error). While these error rates may seem somewhat high. These relatively high error rates may be at least partially due to the conservative artifact rejection. • While our artifact rejection scheme should remove most eye and muscle artifacts, it is possible that some artifact information was not successfully rejected and affected the classification results. Classification Results and Feature Analysis
- 21. Experiment Two
- 22. • The EEG data in this experiment were processed in exactly the same manner as those from Experiment one. • Used exactly the same algorithms for artifact removal, feature extraction, and classification. • Trials with incorrect target selection or reach timeout were excluded. • Any trials showing blink or eye movements were excluded. Before classifying the data. • Result :Reach targets can be successfully decoded from human EEG. Experiment Two
- 23. • This article represents the first steps towards developing BCIs based on EEG recorded during reaching tasks. We described and applied an artifact removal approach suitable for EEG recorded during natural movements. • In Experiment one, we were able to remove movement artifacts from EEG recorded during a natural reaching task and use the cleaned EEG to successfully classify reaching targets. In Experiment two, we decoded reaching targets using short segments of EEG from the premovement planning stage of a delayed reaching task. Conclusion
- 24. • We believe that our conservative artifact removal scheme should be sufficient to remove most movement-related EEG artifacts. This conjecture is supported by good classification performance on our delayed reaching task in Experiment 2, suggesting that our classification results on the natural reaching task in Experiment 1 are due to brain activity and not merely motion artifacts. Conclusion
- 25. • The fact that the cleaned EEG data, especially in the planned reaching task, contained sufficient information to interpret the planned reach targets indicates that: 1) Properly cleaned EEG is a viable data source for studying brain dynamics during reaching. 2) EEG-based BCIs may be able to decode and act on desired reach commands. Conclusion
Editor's Notes
- #18: 1) outer cross-validation loop i = 1:4 a) loop over features j= 1:8 i) extract feature j ii) inner cross-validation loop k = 1:2 A) loop over regularization parameters k = 1:7 iii) select best regularization parameter λ iv) train classifier on feature j with regularization parameter λ b) combine individual classifiers to create meta-classifier i c) compute classification rate on test data i 2) compute average cross-validation classification rate.