GROUP 12

Elimination of eye movement and eye blink based ocular

artifacts from EEG signals
Guided By : Prof. Saptarshi Chatterjee
Team members: Sai Keerthi Namballa [120EI0885]
Abhinash Behera [120EI0702]
Pratikshya Sahoo [120EI0725]
Amrit Nandi [120EI0715]

1
INTRODUCTION

2
The human brain is an intricate and dynamic organ that produces electrical activity,
which can be recorded using Electroencephalography (EEG). EEG signals provide
valuable insights into various cognitive processes and neurological disorders.
However, the interpretation of EEG data is often hindered by artifacts, unwanted
signals that arise from sources other than the brain.

Among the various artifacts, ocular artifacts stemming from eye movements and blinks
are particularly prominent and can significantly impact the accuracy of EEG analysis.
These artifacts introduce noise and distort the genuine neural signals, making it
challenging to draw precise conclusions about brain activity.

3
•EEG Signal Generation: EEG signals are likely composed of synchronous spiking
activities with oscillatory modulation of the local field potential.
•Utility of EEGs: Used to represent brain states (awake, sleeping, selective attention) and
identify active brain regions, particularly on the superior surface of the brain.
•Data Analysis Challenges: Discrimination of signal and noise is challenging due to
various biological signals (e.g., EOG, EMG) and the difficulty in isolating individual
electrophysiological mechanisms.
•EEG in Engineering: Widely used in brain-computer-interface technology.
•Artifact Removal Methods: Traditional methods focus on linear and stationary signal
decomposition. Ocular-related potential is a significant artifact.
•Nonlinear and Nonstationary Properties: Recent attention on signal decomposition
focusing on morphological components due to applicability to nonlinear and nonstationary
signal properties.
•Signal Decomposition Methods: Linear analysis in time-frequency domains (Fourier,
wavelet transforms), blind source separation (ICA, PCA), and sparsity-based methods
(MCA).

4
•Limitations of PCA: Difficulty in reconstructing overall signals due to the ignorance of
signals with small amplitudes and irregular changes.
•Shift to ICA: High order statics in ICA to specify independence in the signal.
•Challenges with ICA: Difficulty in detecting signal components if Gaussian noise is present.
•Plausible EEG Decomposition: Emphasis on sparsity for systematic signal analysis and
synthesis. Introduction of MCA for noise removal based on sparsity.
•Effectiveness of MCA: Demonstrated in image processing, hypothesized to be effective in
EEG artifact removal.
•Proposed EEG Decomposition Method: Based on sparsity and overcomplete dictionaries.
Different dictionaries represent different electrophysiological mechanisms.
•Experiment and Algorithm: Used Block-Coordinate-Relaxation (BCR) algorithm for signal
reconstruction. Goal is systematic artifact removal in EEG signals through MCA.
•Outcome: Highlighted the importance of specifying the best combination of dictionaries for
EEG artifact removal based on EEG frequency properties.

5
Decomposition Using
Morphological Component
Analysis

6
 Recently, decomposition of components in image and time series has a large
expectation in applications, such as minimizing of the data size for transferring the
data via the Internet. MCA based methods fit for the purpose and have the advantage
in the accurate reconstruction of the original data after noise removal, which relies
on the sparsity and over-completeness of the dictionary.

 The concept of sparsity and the over-completeness dictionary has theoretically

extended the traditional signal decomposition to feature extractions focusing on
multiple types of morphologies simultaneously.

 Due to selection freedom of dictionaries, the signal can be decomposed with explicit
dictionary and sometimes it cannot be decomposed in the other form of dictionaries.

7
Key aspects of Morphological Component Analysis:

•It employs multiple dictionaries to represent different morphological components in

the signal.

•It relies on the principle of sparse representation to identify and extract signal
components.

•It offers superior decomposition and reconstruction performance compared to

traditional methods.

•It is particularly well-suited for analyzing signals like EEG data, which exhibit mixed
morphologies.

8
The approximate decomposition of signal S′ into its building components can be expressed as:

Therefore β is the target coefficients for reconstruction of the original EEG signal based on the
assumption ζ ≪ 1, which means that the remainder ζ is negligibly small. In the consideration that ζ represents
the noise part, without noise can be written as

The problem to solve is how optimized coefficients can be

derived, and the equation is rewritten as follows:

9
The problem is how the MCA concept can be embedded in the systems to decompose
biomedical signal especially for EEG signal ?

In this formulation, time-frequency parameters are totally consistent with traditional decomposition
methods which can be applied to the biomedical signal decomposition, such as PCA, wavelets, and
ICA, in the sense of the single set of bases.

One advantage of MCA is the availability of the combination of multiple basis functions, including
traditional basis like wavelet decomposition as a part of the component, called redundant
transforms. Thus, MCA is expected to reveal what kind of the specificity exists in time-frequency
properties of EEG data.
Concrete problems in this viewpoint can be addressed as

(a) what is the best combination of dictionaries of MCA for the EEG decomposition?
(b) what is the true EEG signal in the form of obtained sparsest representation based on selected
dictionaries ϕk?.

10
 MCA decomposition process of arbitrary EEG signal that is assumed to be
a linear combination of k morphological component to be decomposed using
explicit dictionaries.

11
Dictionary Cases

12
Case 1 . —

An overcomplete dictionary ϕ1 is representing the component s1, ϕ1 ∈ MN×L1,

where N ≫ L1, N being the number of samples, that is, the number of time points in the recorded
data.

•For s1, β1opt = arg minβ‖β‖0 subject to s1 = ϕ1β, while solving this equation leads to the sparse
solution (‖β1opt‖0 < ‖β12opt‖0, ‖β13opt‖0).

•For s2, β12opt = arg minβ‖β‖0 subject to s2 = ϕ1β, while solving this equation leads to nonsparse
solution.

•For s3, β13opt = arg minβ‖β‖0 subject to s3 = ϕ1β, while solving this equation also leads to nonsparse
solution.
13
Case 2 . —

An overcomplete dictionary ϕ2 is representing the component s2, ϕ2 ∈ MN×L2,

where N ≫ L2.

•(i) For s2, β2opt = arg minβ‖β‖0 subject to s2 = ϕ2β, while solving this equation leads to the
sparse solution (‖β2opt‖0 < ‖β23opt‖0, ‖β21opt‖0).

•(ii) For s3, β23opt = arg minβ‖β‖0 subject to s3 = ϕ2β, while this equation also has nonsparse
solution.

•(iii) For s1, β21opt = arg minβ‖β‖0 subject to s1 = ϕ2β, while this equation also has nonsparse
solution.

14
Case 3 . —

An overcomplete dictionary ϕ3 is representing the component s3, ϕ3 ∈ MN×L3,

where N ≫ L3.

•(i) For s3, β3opt = arg minβ‖β‖0 subject to s3 = ϕ3β, while solving this equation leads to
the sparse solution (‖β3opt‖0 < ‖β32opt‖0, ‖β31opt‖0).

•(ii) For s2, β32opt = arg minβ‖β‖0 subject to s2 = ϕ3β, while solving this equation leads
to nonsparse solution.

•(iii) For s1, β31opt = arg minβ‖β‖0 subject to s1 = ϕ3β, while solving this equation leads
to nonsparse solution.

15
CODE

16
17
THANK
YOU.