artifact removal from ambulatory eeg
TRANSCRIPT
Md Kafiul Islam
Translational System and Signal Processing Group
Dept. of Electrical & Computer Engineering
National University of Singapore
Purpose:
◦ Diagnosis/Detection of Epilepsy Seizure by Long-term EEG Monitoring (up
to 72 hours)
◦ Early warning of seizures (prediction) onset in order to stop seizure
Special Features
◦ Continuous monitoring
◦ Wireless EEG (up to 10 m or so)
◦ Usually Dry electrode
◦ Few no. of channels
typically 4-8 channels
◦ Patient can lead normal daily life
Except few tasks (e.g. taking bath)
Head Movement ( < 10 Hz )
◦ Shake
◦ Nod
◦ Roll
Hand Movement
Eye Movement
◦ Eyebrow Movement
◦ Eye blinking
◦ Horizontal and Vertical Eye Movement
◦ Extra-ocular muscle activity
Jaw Clenching/Eating/Drinking
Walking
Running
Sleeping
◦ Drowsiness, Light Sleep, Deep Sleep, REM
Talking
Working on PC
Watching TV
And Many More!
Consequence:
Too Many Motion Artifacts!
Challenge: Differentiate
between artifacts, epileptic
seizures and normal EEG activities!
How …!???
Horizontal Eye
Movement (T1-T2)
and
Blink Artifacts (Frontal channels)
EOG
*Chang, B. S., Schachter, S. C., & Schomer, D. L. (2005). Atlas of ambulatory EEG.
Amsterdam ; London: Amsterdam ; London : Elsevier Academic Press, c2005.
Rhythmic eye flutter can occur at quite high frequencies, in this case 5–8 Hz, and trigger automated seizure detection algorithms.
This artifact can be distinguished from ictal events in different ways, including the lack of evolution in frequency or amplitude over time.
Eye Flutter Artifact
recorded by Seizure
Detection Algorithm
Electrode Tapping
Artifact
With ambulatory EEG recording, there may be periods in which the patientis scratching or pressing on the scalp electrodes.
Jaw-Clenching
Artifact
Periods of jaw clenching and biting during ambulatory EEG recording can be associated with artifact.
This activity is often most prominent over the temporal regions, presumably due to temporalis or masseter muscle contraction.
Chewing Artifact
Chewing artifact can often have rhythmic features and at times may resemble ictal activity.
This artifacts are characterized by the rhythmic appearance of muscle artifact centered mostly over the temporal regions bilaterally.
Chewing Artifact
Recorded by
Spike Detection
Algorithm
Chewing artifact can have rhythmic and sharp features, and can thus be picked up by automated detection algorithms.
Here, the spike detections at 23:53:23 and at 23:53:34 capture high frequency sharp waveforms that are artifactual in nature, in association with a nighttime snack.
Dry Electrode Artifact
One disadvantage of ambulatory EEG monitoring is the inability of technologists to check on recording quality frequently in real time and to respond with appropriate technical adjustments as needed throughout the day, as would typically occur in an inpatient monitoring unit.
Here, a dry electrode at F8 causes an artifact seen throughout this excerpt.
Forehead Rubbing
Artifact
As with many other patient movement artifacts, rhythmic rubbing artifact can be a potentially confusing finding on ambulatory EEG monitoring, particularly in the absence of concomitant video recording. Here, rhythmic rubbing of the forehead produces an artifact in the anterior channels, extending from the beginning of this excerpt through 20:50:16.
Differentiation of such artifact from an ictal event, based on the lack of evolution of the artifact in frequency or amplitude, is critical.
Pulse Artifact
Pulse artifact is caused by the rhythmic movement of scalp electrodes, usually over the temporal regions, by the pulsations of blood through vessels close to the skin, such as the superficial temporal artery. It is to be distinguished from the more common electrocardiographic (EKG) artifact in that pulse artifact is typically not as sharply contoured, resembles a slow wave, and follows the QRS complex rather than being simultaneous to it. Here, pulse artifact is seen in channels 9 and 10 over the left temporal region throughout this entire excerpt.
Delta: (< 4 Hz)
Theta: (4 - 8 Hz)
Alpha: (8 - 13 Hz)
Beta: (14 - 30 Hz)
Gamma: (> 30 Hz)
Mu: (7.5 - 12.5 Hz)
Gamma
Use of an extra sensor
◦ Adaptive filtering
Accelerometer
Gyroscope
Contact Impedance Measurement
A-priori user input required
◦ Wiener/Kalman/Particle Filtering
Blind Source Separation
◦ ICA/CCA/MCA
Time-series Analysis
◦ Wavelet transform / STFT
Empirical Technique
◦ HHT/EMD/EEMD
Artifact Generation Protocol (From Literature)
Domain◦ Time◦ Frequency◦ Wavelet
Statistical Features◦ Entropy
ApEn, MSE, CMSE, MMSE, etc.
◦ Kurtosis◦ Skew◦ RMS/Variance/S.D.◦ Max/Min/Mean◦ Maxima/Minima
High Frequency Oscillations (HFO)80 Hz – 200 Hz
Ripple200 Hz – 600 Hz
Typical Scalp EEG!0.05Hz – 128 Hz
Seizure Appears in Scalp EEG0.5 Hz - 29 Hz [1,2]
1. J. Gotman, “Automatic recognition of epileptic seizures in the EEG,”Electoencephalogr Clin. Neurophysiol., vol. 54, pp. 530–540, 1982
2. M. E. Saab and J. Gotman, “A system to detect the onset of epilepticseizures in scalp EEG,” Clin. Neurophysiol., vol. 116, pp. 427–442,2005.
Seizure Appears between 0.5 Hz to 29 Hz
Raw EEG
D1
(64 – 128 Hz)
A1
(0 – 64 Hz)
D2
(32 – 64 Hz)
A2
(0 – 32 Hz)
D3
(16 – 32 Hz)
A3
(0 – 16 Hz)
D4
(8 – 16 Hz)
A4
(0 – 8 Hz)
D5
(4 –8 Hz)
A5
(0 – 4 Hz)
D6
(2 – 4 Hz)
A6
(0 – 2 Hz)
D7
(1 – 2 Hz)
A7
(0 – 1 Hz)
D8
(0.5 – 1 Hz)
A8
(0 – 0.5 Hz)
Multilevel Wavelet DecompositionFs = 256 Hz, L = 8, Haar Wavelet
Artifact Removal Applied on Simulated Data
o Simulated Clean EEG (downloaded from [1])
o Simulated Seizure (downloaded from [2])
o Artifact Simulation (to mimic extreme ambulatory condition)
[1]. http://www.cs.bris.ac.uk/~rafal/phasereset/[2]. Newborn EEG Seizure Simulation
Performance Evaluation of Proposed Artifact Removal
-0.8 -0.7 -0.6 -0.520
25
30
35
40
45
50
55
Artifact SNR in dB
%ar
tifac
t red
ucti
on, l
amda
o The target is to remove artifacts as much as possible
o Artifact removal should enhance the performance of later signal processing stage (e.g. increase seizure detection rate)
o Average amount of artifact reduction is around 35% when artifact SNR is around 0 dB
o Not all artifacts affect seizure detection performance
EEG Features before and after Artifact Removal
Features Extracted: (i) Entropy (ii) Kurtosis (iii) Line Length (iv) Peak(v) NEO (vi) Variance (vii) FFT (viii) FFT Peak
Note: The features between seizure and non-seizure data are more separable after artifact removal which suggests that it increases the detection rate and minimizes false alarms (false alarms are due to artifacts).
Characterize movement artifacts by doing EEG experiments
Use real seizure data marked by the epilepsy specialists
Improve the algorithm by avoiding ICA
Use classification techniques to differentiate seizure from non-seizure data and evaluate the performance of proposed algorithm