Please enable JavaScript.
Coggle requires JavaScript to display documents.
Brain-Computer Interfacing - Coggle Diagram
Brain-Computer Interfacing
Machine Learning
Unsupervised (Weight matrix)
Principal Component Analysis (PCA)
Independent Component Analysis (ICA)
Supervised (Weight vector)
Linear Regression
Linear Discriminant Analysis
Logistic Regression
ERP decoding pipeline
Part 1: Loading and preprocessing
Bandpass + downsample (Nyquist) Inspect and remove outlier channels Inspect and remove outlier subspace via ICA
Part 2: Epoching and feature extraction
Windowing into epochs and remove outliers Extract average potential of a few time intervals
Part 3: Training of decoding model
Train LDA or logistic regression Estimate generalization error
Part 4: Online Application
Record ongoing brain signals Preprocess (go to Part 1) and Extract features (go to Part 2) Classify features with model Provide feedback to user Dynamic Stopping
Model Evaluation
Confusion Matrix
True Positive Rate - Sensitivity
True Negative Rate - Specificity
Accuracy/Error rate for binary classification
Area under the receiver-operating characteristic curve
Information Transfer Rate
User Experience/Usability
(fold) Cross-Validation
Oscillatory Pipeline
Typical Oscillation
Beta rhythm (13-30 Hz)
:
Active thinking, motor activity
Alpha rhythm (8-13 Hz)
: Awake relaxation
Gamma rhythm (>35 Hz)
: Focused attention
Theta rhythm (4-8 Hz)
: REM sleep, voluntary control of movement
Delta rhythm (0.5-4 Hz)
: Deep REM sleep, memory consolidation
Envelopes
Transient and from spindles
Calculation via Hilbert Transform or approximation by power/variance
α Event-Related Desynchronization
: upon start of motor task, typically contra-lateral
α Event-Related Synchronization
: towards end of motor task
β Event-Related Synchronization
: rebound at the end of (brisk) motor task
(supervised) Feature extraction
Common Spatial Patterns
Decomposes original channel space linearly into subspace components
Represents each oscillatory subspace by a spatial filter
Contrast enhancement --> A weight contained in spatial filter
w
either enhances a channel's contribution to the mixture, or suppresses
Selecting components showing the most extreme eigenvalues
Source Power Comodulation
Can we find an oscillatory component which comodulates with the external stimulus?
SPoC tries to derive a spatial filter
w
which maimizes correlation and covariance between the bandpower of an estimated source and the external stimulus
Choose the filters with the strongest eigenvalues to derive informative SPoC components.
Linear Regression
Use labelled data points to train a model which can predict the label y of a new data point x
Learn a function
f(x)
, which shall predict the label y based on a data point x
y=Xw+ϵ
Assumptions required for analytical solution
The residual errors are uncorrelated and share the same variance
The residual errors follow a normal distribution
The expected value of the residual errors is zero
Analytical solution:
