BCI background

If you ever wondered the world of Brain-Computer Interface, we are in the same place. So interested I was in it, that I decided to prepare a year project related to BCI.
In order to accomplish the proposed project, I had to do an extensive research on BCI. It was undoubtedly a time-consuming task,but also rewarding at the same time. The knowledge a I gained supported to build my own draft of an integral framework of the BCI world.
I would really want to develop each of the points into proper paragraphs linking each logically. However, I decided to not disclose them fully until the full research is finished. So for now, I am presenting a series of words which are concepts that can help you to constitute your own BCI mind map. In case you need a further reading, I have also attached some further reading which were most helpful for me.

Non-invasive vs Invasive

EEG preferred, 83% of BCI are EEG.

Paradigms

Setups and target of study.

First generation

Sensorimotor rhythms
P300
Visual Evoked Potential
Slow Cortical Potential
Activity of Neural Cell

Second expansion

Added:

Response to mental task
Hybrid

The most studied paradigms now

Motor Imagery
- Sensorimotor rhythms SMR
- Imagined Body Kinematics IBK
External Stimulation
- P300
- Steady State Visual Evoked Potential SSVEP
Error-potential
Hybrid/others

Mechanisms of Motor Imagery

Mu rhythm and beta rhythms
Normally brain in ERSynchronisation
When trying to execute an action -> ERDesynchronisation
Mainly in frontal and medial lobe
- C3, C4 most relevant
Capture these changes, identify patterns, predict actions

Methodology in Motor Imagery

Pipeline:
1. Collect amplitude
2. BCI Transducer
  1. Artifact Processor
  2. Feature Generator
    1. Signal enhancement / Feature pre-processing
    2. Feature extraction (paradigm dependent)
    3. Feature selection / Dimension reduction
  3. Feature Translator
    1. Feature classification
    2. Post-processing
3. Control Interface
4. Device Controller
A catalogue of methods (till 2007, review 1)
- See Table 1 in Appendices
A review 2 on methods till 2007
- Feature filtering methods
  - Time domain
  - Spatial domain
- Feature extraction methods
  - Band-power features
  - Time point features
  - Connective features
- Feature selection
  - Filter
  - Wrapper
  - Embedded methods
- Performance measurer
  - Classification accuracy
  - Confusion Matrix or Kappa metric
  - Receiver Operating Characteristics or Area Under the Curve
- Classifiers: 5 families
  - Linear classifier
  - Neural Networks
  - Non-linear Bayesian classifiers
  - Nearest Neighbour classifiers
  - Combinations
A review 3 up to 2018
- Most frequent methods
  - Band-pass filtering
  - Common Spatial Patterns
  - Linear Discriminant Analysis
Review 2’s update till 2017, future expectations, second generation methods
- Adaptive classifiers
  - Preferred, outperforms, regardless supervised or not
- Matrix and tensor classifiers
  - Promising, requires more research
  - Riemannian geometry classifier is at the state-of-art
- Transfer learning methods
  - Useful when few data
  - Performance highly variable
- Deep learning
  - Not convincing results
  - May due to limitation of data
  - Shallow CNN may work better
- Miscellaneous but surprisingly good simple models when little data
  - Shrinkage LDA (always over LDA)
  - Random Forest classifiers

Performance of Motor Imagery paradigms, a review

Importance of personal information + psychological + physiological + anatomical variables.

Most frequent setup

Number of channels
Number of states
Number of subjects
…
Classification Accuracy
Overall Corrected CA in the literature
Correlations between setup variables and accuracy
Few effective control

BCI Challenges

BCI Illiteracy is causing performance variation

Drives to ERD illiteracy. Difficult the transferability of a valid BCI. Inter-subject BCI is still difficult.

Online vs Offline

Most of them are offline, while online is the important part.

Advanced techniques still to be explored

Incorporate ANN, adaptive classifiers, Riemmanian Geometry Classifier, etc.

Lack of standardised manner of summarising key features

Resources

The following are titles of some papers that I found useful to understand the BCI world.

A comprehensive review of EEG-based brain–computer interface paradigms
Most Popular Signal Processing Methods in Motor-Imagery BCI: A Review and Meta-Analysis
A survey of signal processing algorithms in brain–computer interfaces based on electrical brain signals
A review of classification algorithms for EEG-based brain–computer interfaces: a 10 year update
Chapter 2 - Technological Basics of EEG Recording and Operation of Apparatus
Oscillatory gamma activity in humans and its role in object representation
Could Anyone Use a BCI?
Performance variation in motor imagery brain–computer interface: A brief review

Appendices

Table 1

Stage	Methods	Subtypes	Percentage
Pre-processing methods 96	Surface Laplacian		32%
	Principal Component Analysis or Independent CA		22%
	Common Spatial Pattern		14%
	Common Average Referencing		11%
	Frequency Normalization, Common Spatial Subspace Decomposition, Robust Kalman Filter, Singular Value Decomposition, Local Averaging Technique, Wiener Filtering, Sparse Component Analysis, Maximum Noise Fraction, Spike Detection Methods, Neuron Ranking methods, Common Spatio-Spatial Patterns
Feature Selection methods 38	Genetic Algorithm		26%
	Distinctive Sensitive Learning Vector Quantization		24%
	PCA		13%
	Sequential Forward Feature Selection, Grid Search Method, Relief Method, Recursive Feature/channel Elimination (RFE), SVM-based RFE, Stepwise Discriminant Procedure, Linear Discriminant Analysis (LDA), Fisher LDA (dimensionality reduction), Fisher Discriminant-based Criterion (feature selection), Zero-norm Optimization, Orthogonal Least Square based Radial Basis Function
Post-processing methods 30	Averaging and thresholding		57%
	Debounce Block to deactivate		27%
	Event-related Negativity		16%
Feature Extraction methods	Sensorimotor	Power-Spectral-Density	41%
		Parametric modelling (Autoregressive)	16%
		Time-Frequency Representation	13%
		CCTM, Signal Envelope, Hjorth Parameters, Signal Complexity, Lateralized Readiness Potential features, others
		Amplitude (None)	6%
	VEP	Power-spectral features	64%
		Lock-in Amplifier, Asymmetry Ration of Different Band Powers, Cross-correlation, Amplitude between N2 and P2 peaks, none
	P300	Peaks of specific time window	26%
		TFR-based methods	22%
		Cross-correlation	15%
		Stepwise Discriminant Analysis, Matched Filtering, Piecewise Prony Method, Area Calculation, None
	Mental task	Power-Spectral features	41%
		Parametric modelling (Autoregressive)	37%
		Signal Complexity, Eigenvalues Correlation Matrix, Linear Predictive Coding using Burg’s Method
	SCP	Low-pass filtering and amplitude	74%
		Time Frequency Representation (TFR) method, Mixed Filter, None
Feature Classification methods	Use of Threshold detector		27%
	Linear Discriminant, LDA or FLD		26%
	Neural Networks: 27% are Multi-Layer Perceptron		25%
	Continues Feedback		16%
	SVM		11%
	Hidden Markov Model, Bayesian Classifier (Linear), Gaussian Classifier, Signal Space Projection (SSP), Mahalanobis Distance based classifier, Nonlinear Discriminant Function, Bayes Quadratic Classifier, Linear Bayesian Decision Rule, Logistic Regression, Z-scale-based Discriminant Analysis, Linear Dynamical System, Self-Organized Map based SSP, KNN, Variational Kalman Filter, Random Forest Algorithm

Core Concepts in BCI Literature

In my way to a Brain-Computer Interface project, I did an extensive research on this field. This post then enumerates sets of key terms which formed my basis of understanding to BCI literature.