Professional Documents
Culture Documents
PL Artifact Detection Removal Analysis Report
PL Artifact Detection Removal Analysis Report
● Categories:
1. Frequency domain methods: assumes not all EOG Artifact signals effect scalp sample sites
■ Leads to exclusion of data segments
2. Source localization methods: attempts to quantify effect to which artifact would propagate
to scalp sites
■ ICA, BSS
■ Editing of segments via removal of artifact
● Assumptions:
○ Eye artifacts and EEG are independent
○ EOG and EEG are bidirectional interference
● Data format:
○ Rows - EEG channels
○ Columns - time point measurements
Distortion
1. Undercorrection - If artifacts are
missing in detection phase,
artifacts will remain in data
2. Overcorrection - If artifacts and
noise free signals are detected
and removed or if data does not
include artifacts
No Free Lunch Theorem
● Optimization approach
● Functionality: Separates linearly mixed sources into independent components (IC’s) via linear
regression
○ Preforms independent component decomposition and analysis
○ Finds noise free EEG IC and artifact IC
○ Reduces artifact IC weights to 0
■ = Nullifies/disregards artifact IC’s
○ Uses regluationzation and normalization to reduces weights and computational complexity
○ Minimizes Gaussianity of projection on all axises (‘Whitening”)
● Goal: find maximal independent components
○ Determine weights (describe noise free EEG and EOG)
■ X estimate = W*S
● S = contaminated original EEG data set
● X = IC’s
■ Linear model of S = M*X where M is the linear weight of IC’s and X are IC’s
ICA Continued
● Requirements:
○ Whitening (preprocessing of data)
■ Remove correlations
■ Linear change of coordinate
■ Rotates data and makes sphere like shape in scatter plot
○ Large amounts of data
● Assumptions:
○ Signal sources of noise free EEG and artifact are instantaneously linear mixtures that can be
decomposed into IC’s
○ Number of components < = number of channels
○ Summation of signal sources are linear
○ Artifacts and noise free EEG are spatially fixed across time
○ Sources are non Gaussian
ICA Continued
● Requirements:
○ Data must include Exogenous reference channels - Electrooculogram (EOG)
● Pros:
○ Does not have number of channels constraints (can use minimum without effect)
Linear Regression Continued
● Goal: Find and define relation between noise free EEG and artifact signals
○ Find regression coefficients (weight matrix) that describe the correlation between EOG artifact and
noise free EEG channel
■ E = S - WB
● S = Original mixed EEG signal
● E = noise free EEg w/out artifact
● B = artifact
● W = weight matrix
Other Methods
● Tools:
○ Raw data review
○ Data pre processing
○ Artifact rejection and correction
○ Data averaging and ERP analysis
○ 2D and 3d Mapping
○ Source Analysis
○ Time frequency analysis
● Run independently of EEG and MEG obtainiment
● Analysis software w/ direct MatLab interface connection (Path)
● Has artifact correction Module
BESA Artifact Correction Module
● Automated EOG Artifact ICA algorithm based on spatial distribution of scalp maps
○ Included in Measurae Projection Toolbox (MPT)
○ Performs ICA with focus on scalp map features
○ Comparable to CORRMAP
■ Found to have .993 area under ROC curve: “Demonstrating it has high sensitivity and
specificity” [9].
○ Compares input to an EOG IC scalp map database (has more than .5 million samples)
EEGLAB
● Manually:
○ User identifies segments that are obstructed via artifact
○ Results in list of segments w/ artifacts
● Automatic:
○ “Although automatic artifact detection algorithms work efficiently for well-known artifacts in
well-behaved data, you should not use automatic detection functions as your defaults method”
[6].
○ Composed of filtering and channeling combination techniques
■ Various linear and non linear filtering
■ Data padding options
○ Functions:
■ ft_artifact_clip, ft_artifact_ecg, ft_artifact_threhosld, ft_artifact_jump, ft_artifact_muscle,
ft_artifact_avalue
○ Offer PCA, ICA, etc.
MNE Python
● There are more citations for Matlab EEGLab and FieldTrip in comparison to MNE Python [2]
Future:
● Questions:
○ How EEG data was preprocessed?
■ Manually or w/ semi-automatic artifact detection methods?
○ Was there use of EOG electrodes?