PhD Oral Defense ARTIFACT CHARACTERIZATION, DETECTION AND REMOVAL FROM NEURAL SIGNALS Presented By: Md Kafiul Islam (A0080155M) Supervisor: Dr. Zhi Yang Department of Electrical and Computer Engineering National University of Singapore 28 th Oct, 2015
48
Embed
PhD Oral Defense of Md Kafiul Islam on "ARTIFACT CHARACTERIZATION, DETECTION AND REMOVAL FROM NEURAL SIGNALS"
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
PhD Oral Defense
ARTIFACT CHARACTERIZATION, DETECTION AND REMOVAL FROM NEURAL SIGNALS
Presented By: Md Kafiul Islam
(A0080155M)
Supervisor: Dr. Zhi Yang
Department of Electrical and Computer Engineering National University of Singapore
28th Oct, 2015
Outline
• Background • Problems and Motivation • Thesis Objectives • Literature Review • Presentation of Thesis Contributions
– Artifact Study on in-vivo neural data – Proposed Artifact Removal Algorithms
• In-Vivo Neural Signals • EEG for Seizure Detection and BCI
EEG is the recording of the brain's spontaneous electrical activity over a period of time by placing flat metal discs (electrodes) attached to the scalp.
• EEG Rhythms
• Transients
Background-2: EEG and its Characteristics
Scalp EEG is Most popular and widely used brain recording technique
1) Low-cost 2) Non-invasive 3) Easy to use 4) fine temporal resolution
Typical Scalp EEG B.W.: 0.05Hz – 128 Hz
4
Motivation-1
Artifacts are unwanted signals originated from non-neural
source
Recordings corrupted by artifacts, especially in less constrained
environment.
Cause mistakes in interpretation of neural information.
Artifacts need to be identified and removed for reliable data
analysis.
The challenges for in-vivo artifact identification compare to EEG
artifacts are:
No prior knowledge about artifacts unlike EEG-artifacts
The broad frequency band of in-vivo data (0.1 Hz – 5 kHz)
makes it difficult to separate artifacts from signal
Existing artifact removal methods are intended for EEG, So can’t be
• To investigate artifacts present at in-vivo neural recordings: characterize them and observe
the change in dynamic range.
• To propose an automated artifact detection and removal algorithm for reliably remove artifacts from in-vivo neural recordings without distorting signal of interest
• To synthesize an artifact database for quantitative performance evaluation of any artifact removal method.
• To propose application-specific artifact removal methods for scalp EEG recordings • Epilepsy seizure monitoring and detection purpose
• BCI studies/experiment purpose
• To observe the after-effect of artifact removal on later-stage neural signal processing. i.e. • Improvement in neural spike detection (in-vivo)
• Improvement in epileptic seizure detection (EEG)
• Improvement in BCI classification (EEG)
8
Literature Review (No literature particularly on artifacts for in-vivo neural signals)
Can work for both single and multi-channel recordings
Most importantly it can be used for both detection (from decomposed
coefficient) and removal (thresholding and reconstruction) of artifacts.
Why SWT Preferred over DWT or CWT?
Usually DWT or SWT is preferred over CWT when signal synthesis is required
CWT is very slow and generates way too much of data.
SWT is translation invariant where DWT is not. So better reconstruction result (No loss of information, preserves spike data and doesn’t generate any spike-like artifacts).
Choice of mother wavelets for CWT is limited.
SWT implementation complexity [O(N L)] is in between DWT [O(N)] and CWT [O(N L log2N)].
N = length of signal, L = decomposition level
Digital implementation of SWT: A 3 level SWT filter bank and SWT filters
k = kA for approx. coef. kD for detail coef. By empirical observation from signal histogram 5 < m < infinite 2 < n < 3 D3, D4, D5, D6 => Spikes. D8, D9, D10 and A10 => LFP
21
Choice of Threshold Function (Garrote) • Hard: Discontinuous which may produce large variance (very sensitive to small changes
in the input data)
• Soft: Continuous but has larger bias in the estimated signal (results in larger errors)
• Garrote: Less sensitive to input change, lower bias and more importantly continuous.
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).
Improvement in Seizure Detection (Cont…)
Algorithm Design-3: Artifact Detection and Removal from EEG for BCI
Scalp EEG-based BCI is the most widely used BCI studies 1. P300 ERP (Event Related Potential)
2. MI (Motor Imaginary)
3. SSVEP (Steady-state Visual Evoked Potential)
Challenges
Difficult to avoid artifacts during BCI experiments
Approaches – Unique idea of Artifact Probability Mapping
– Epoch by epoch processing
– SWT-based denoising
– Consideration of type of BCI to utilize desired signal band(s) for artifact identification.
Comparison of Current EEG Artifact Removal Techniques With Proposed Ones
EEG Artifact Removal for Seizure Detection EEG Artifact Removal for BCI
Co
mp
uta
tio
nal
Tim
e Pe
rfo
rman
ce M
etri
cs V
alu
e
40
Summary of Contributions
• Investigation on In-Vivo Neural Artifacts (for the very First Time) – Identifying artifact sources – Characterizing them in to 4 types – Studied change in dynamic range
• Artifact Database Synthesis – Allowing realistic artifact simulation in real clean neural signals – Quantitative performance evaluation becomes possible
• Unique Artifact Probability Mapping – Gives user the freedom to select probability threshold – Applicable to other EEG applications
• Proposed 3 different artifact removal algorithms (First time for in-vivo neural data)
– Almost no distortion to neural signal of interest – Doesn’t depend on artifact types – Application specific solution – Can work for both single and multi-channel neural data – Parameters can be optimized for best performance – Straightforward parameter adjustment. – Automatic algorithm / Minimal manual intervention (during initial training
parameters) – Suitable for both online and offline processing – Unique idea of artifacts probability mapping for EEG epochs – All three algorithms’ performances have been evaluated both qualitatively and
quantitatively. – Compared with other existing competing methods and ours found to be
superior – Open source codes available for everyone to use and edit for further