Inter-electrode correlation (IEC) algorithm for reducing common-noise artifacts in multichannel neural recordings (IEEE/EMBS 2008)

1. Automated Reduction of Non-neuronal
Signals from Intra-cortical Microwire
Array Recordings by Use of
Correlation Technique
Kunal J. Paralikar1
, Chinmay Rao2
, and Ryan S. Clement1
8/21/08
30th
Annual International Conference of the
IEEE Engineering in Medicine and Biology Society
1
Department of Bioengineering and 2
Department of Electrical Engineering
The Pennsylvania State University

2. Outline
• Statement of Problem: common noise
• Basic Approach
– Inter-electrode Correlation (IEC) algorithm
– Performance metrics: Mean-spike
• Initial Evaluation Study
• Comparison Study
– Differential ref, virtual ref, IEC w/ varying levels of
common noise
• Conclusions

3. A
B
0
1
2
3
4
5
6
R44 R46* R47 R48 R50 R51
Subject
MaxForce(mN)
Collagenase Sites
Control Sites
coll Non-coll
Average
p < 0.015
A B
Days Post-Implant

4. 0 2 4 6 8 10 12 14 16 18 20
-80
-60
-40
-20
0
20
40
60
uVolts
msec
Model-based analysis of cortical recording with silicon
Microelectrodes (Michael A. Moffitt, Cameron C. McIntyre*)
Intracortical Microelectrode Recordings
Univ. of Utah

5. Classic Signal Processing Chain
Spike detection
threshold
Filtered signal
PCA Clusters Template Sorting
Detect
Spikes
Classify/
sort
Analyze
Spiketrains
Detected
spikes
Classified Spikes
Figures from:
www.plexoninc.com

6. 0 2 4
Time (s)
200
0
-200
Multi-channel Recording Sample
Voltage(µV)
Common Noise- A Common Problem
Common Noise Sources:
- Movement artifacts
- Electric fields
- Physiologic (LFPs, EEG, EMG)

7. -200
0
200
1300 65
Time (ms)
-60
-80
40
40
30
Time (ms)
Voltage(µV)
Voltage(µV)Voltage(µV)
Two Types of Suprathreshold Spiking Events
E1
E2
E3
E4
Spike on E3
Segments on other electrodes

8. Omnetics
connector
dental acrylic
skull bone
cortex
(motor area)
1mm
50 µm
250 µm
250 µm
polyimide-coated
tungsten
connector
interface board
Omnetics
connector
dental acrylic
skull bone
cortex
(motor area)
1mm
50 µm
250 µm
250 µm
polyimide-coated
tungsten
connector
interface board
Model-based analysis of cortical recording with silicon
Microelectrodes (Michael A. Moffitt, Cameron C. McIntyre*)
Distant vs. Local Signal Sources
An extracellular action potential from a local neuron is not
likely to register the same waveform on multiple wires

9. F
PCA Analysis
-80
-40
0
40
Voltage(µV) Correlated segments only
Uncorrelated segments
only
E
Time (ms)
0 3
Mean Spike Analysis
0 2000 4000 6000
4000
8000
Frequency (Hz)
Threshold Detected
Uncorrelated
Only Correlated
D
Spectrum Analysis
F
PCA Analysis
Common-noise spikes often look neural !
…simple spike detection and sorting based on
individual electrodes is limited.
Can we use other electrodes to help?
Individual Electrode Analysis Limitations

10. Inter-electrode Correlation
(IEC) Algorithm
• Design Goals:
– Reduce false-positive spike detection by
identifying and removing highly correlated
spiking events
– Be conducive to quantitative assessment of
neural interface recording performance (no
altering of signals at their source)
– Simple and objective

11. 0 5
Time (s)
200
-200
Voltage(µV)
E1
E2
E3
E8
Amplification/filtering, A/D conversion, digital filtering
analyzed
electrode
Experimental Overview: Signal Processing
Spike
Detection
*
Correlation
Algorithm
Mean Spike
(T-PCA:
Threshold+PCA)
Mean Spike
(T:Threshold
Only)
Mean Spike
(Corr)
***
Mean Spike
(Corr-PCA)
*
*
Principal
Component
Analysis
Principal
Component
Analysis
Omnetics
connector
dental acrylic
skull bone
cortex
(motor area)
1mm
50 µm
250 µm
250 µm
polyimide-coated
tungsten
connector
interface board
Threshold-based Spike
Detection Block
*
Voltage(µV)
40
-80
0
***
0 0.75 1.5 2.25 3
-100
0
Peak to Peak
Amplitude
Depolarization
Phase
Repolarization
Phase
Voltage(µV)
60
time (ms)
All Valid Spikes
Mean Spike
Mean Spike CalculationDetails of Correlation Algorithm
Time (ms)
0 31.5
40
-50
0
0 31.5
Voltage(µV)
Spike with low
inter-electrode
correlation
Spike with high
inter-electrode
correlation
KEEP! DISCARD!
** R<0.75 R>0.75

12. -100
60
0
uV
2.251.50.750 3
ms
Rater Selected
unsorted
80
90
100
110
120
130
140
150
410 420 430 440 450 460 470 480
Depolarization Phase (uSec)
PeaktoPeakAmplitude(uV)
E1
E4
U
U
RS
RS
Rationale for Performance Evaluation
Increasing neural content
Assumption:
Mean-spikes generated by true
neural spikes will be sharper
and larger in amplitude
Unsorted= includes all spikes that crossed threshold
Rater Selected= manually selected by experienced user

13. Data Set Details
• Neural recordings from our previous collagenase-aided
electrode insertion study*
• Chronic microwire implants (8ch) in rat motor cortex
• 12kHz sampling (300-5kHz filtering)
• 1 data set = 5 minutes of recording obtained during a
daily recording session (10 total sets)
• Subjects were awake during recording
* Paralikar, K.J. & Clement, R.S. Collagenase-Aided Intracortical Microelectrode Array Insertion: Effects
on Insertion Force and Recording Performance. IEEE Transactions on Biomedical Engineering. Vol 55,
No. 9, September 2008. (Available Online)

14. 2.25 30 0.75 1.5
ms
2.251.50.750
-70
40
0
uV
Thresholding and Correlation
Thresholding Alone
Results: Mean-Spike Waveform Samples
-600 -400 -200 0 200 400 600
-600
-400
-200
0
200
400
600
Center 10,10
Radius 200
-600 -400 -200 0 200 400 600
-600
-400
-200
0
200
400
600
Center 10, 10
Radius 200
PCA Analysis
Threshold Alone Threshold w/ Corr
2.251.50.750
-70
40
0
ms
uV

15. Duration(µs)
Duration(µs)Voltage(µV)
Depolarization Phase Repolarization Phase
Peak-Peak Amplitude
0 0.75 1.5 2.25 3
-100
0
Peak-Peak
Amplitude
Depolarization
Phase
Repolarization
Phase
Voltage(µV)
60
time (ms)
N = 10 electrodes
* p<0.01
T: Thresholded Only T_PCA: Thresholded w/ PCA
Corr: Threshold + Correlation Corr_PCA: Thresh + Correl + PCA
Results: Mean-Spike Feature Analysis
* *
* *

16. NumberofEvents
Results: Event Detection Rates
• Use of inter-electrode correlation significantly reduces event counts.
• This could lessen down the load on downstream processes.
** **
**p<0.001

17. Initial Performance Conclusions
• Removing highly correlated spike events
improves the mean-spike shape (more
neural-like)
• Inter-electrode correlation is a simple
technique to reduce false positive spikes

18. Comparative Study*
• Study Goals:
– Compare results from:
• Simple-threshold
• Differential referencing
• Virtual reference (common average subtraction)
• Inter-electrode correlation
– Explore performance under different degrees
of common noise
*Paralikar K, Rao C, Clement R. New approaches to eliminating non-neuronal
artifacts in recordings from intra-cortical microelectrode arrays: inter-electrode
correlation and virtual referencing. Journal of Neuroscience Methods
(Submitted)

19. Data Set SelectionData Selection
0
20
40
60
80
100
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
Data Set
PercentageDifferenceor
CorrelationCoefficient
RMS difference b/w electrode and virtual reference
Correlation Coefficient
Low Common Noise Medium Common Noise
High Common Noise
• Each data set was composed of 5-minutes of neural recordings from motor
cortex of chronically implanted rats

20. 0 5
Time (s)
200
-200
Voltage(µV)
E1
E2
E3
E8
Omnetic
s
connect
or
dental
acrylic
skull bone
cortex
(motor area)
1mm
50
µm
250 µm
250 µm
polyimide-coated
tungsten
connector
interface
board
Amplification/filtering, A/D conversion, digital filtering
Ensemble Average
“Virtual Reference”
differential
reference
analyzed
electrode
Experimental Overview: Signal Processing
Spike
Detection
Spike
Detection
Spike
Detection
Correlation
Algorithm
Mean Spike
(Differential)
Mean Spike
(Threshold Only)
Mean Spike
(Correlation)
Mean Spike
(Virtual Reference)
Standard Thresholding Virtual Referencing
Inter-electrode CorrelationDifferential Reference
Signal Processing Paths:

21. Results: Mean-spike waveform samples
Mean Spikes
Correlation = 0.2
-120
-80
-40
0
40
Threshold detected
Virtual Referencing
Differential Recording
Inter-electrode Correlation
Low Common Noise Example
0 3
Voltage(µV)
Time (ms)
3
Correlation = 0.5
Medium Common Noise Example
0
Threshold detected
Virtual Referencing
Differential Recording
Inter-electrode Correlation
Correlation = 0.8
High Common Noise Example
0 3
Threshold detected
Virtual Referencing
Differential Recording
Inter-electrode Correlation
Mean Spikes
Correlation = 0.2
-120
-80
-40
0
40
Threshold detected
Virtual Referencing
Differential Recording
Inter-electrode Correlation
Low Common Noise Example
0 3
Voltage(µV)
Time (ms)
3
Correlation = 0.5
Medium Common Noise Example
0
Threshold detected
Virtual Referencing
Differential Recording
Inter-electrode Correlation
Correlation = 0.8
High Common Noise Example
0 3
Threshold detected
Virtual Referencing
Differential Recording
Inter-electrode Correlation
Correlation = 0.2
-120
-80
-40
0
40
Threshold detected
Virtual Referencing
Differential Recording
Inter-electrode Correlation
Low Common Noise Example
0 3
Voltage(µV)
Correlation = 0.2
-120
-80
-40
0
40
Threshold detected
Virtual Referencing
Differential Recording
Inter-electrode Correlation
Threshold detected
Virtual Referencing
Differential Recording
Inter-electrode Correlation
Low Common Noise Example
0 3
Voltage(µV)
Time (ms)
3
Correlation = 0.5
Medium Common Noise Example
0
Threshold detected
Virtual Referencing
Differential Recording
Inter-electrode Correlation
3
Correlation = 0.5
Medium Common Noise Example
0
Threshold detected
Virtual Referencing
Differential Recording
Inter-electrode Correlation
Threshold detected
Virtual Referencing
Differential Recording
Inter-electrode Correlation
Correlation = 0.8
High Common Noise Example
0 3
Threshold detected
Virtual Referencing
Differential Recording
Inter-electrode Correlation
Correlation = 0.8
High Common Noise Example
0 3
Threshold detected
Virtual Referencing
Differential Recording
Inter-electrode Correlation
Threshold detected
Virtual Referencing
Differential Recording
Inter-electrode Correlation
Inter-electrode correlation method (- - -) consistently produced the
most neural appearing mean-spike waveforms under all common
noise conditions

22. Results: Mean-spike feature analysis
Medium Common-noiseLow Common-noise High Common-noise
Depolarization Phase Duration (µs)
Mean-spikePeak-to-Peak(µV)

23. 0
0.2
0.4
0.6
0.8
1
1.2
1.4
Low Medium High
EventCounts(normalized)
Relative Common-noise Level
Thres
VirtualRef
Differential
Correlation
Results: Event Detection Rates
Low Common Noise
Thres-Detected Virtual-Ref Differential Inter-Elect Cor
Dep Phase (uSec) 461.3±18 475.6±19.8** 796.8±212.4** 465.8±17**
Rep Phase (uSec) 1401.7±268.6 1304±183.5 883.6±542.9** 1254.2±131.7
Pk to Pk Amp (uV) 110.6±19.8 91.5±17.7** 121.2±17.2 100±12.6
SpikeSegments 11055±2432 11211±1893 9957±1501 8791±1914**
Med. Common Noise
Dep Phase (uSec) 641.9±197.4 515±1467.1 575±72.8 497.6±49.2
Rep Phase (uSec) 1357.5±410.5 1467.1±63.4 1355.9±205.3 1359.2±79.6
Pk to Pk Amp (uV) 98.2±14.1 70.4±12.2** 87.6±15.5 93±10
SpikeSegments 7713±1982 9513±2760** 8088±1979 4407±2688**
High Common Noise
Dep Phase (uSec) 760.3±214.3 514.3±85.9 615.6±124.7 544.6±81.9
Rep Phase (uSec) 1083±662.4 1628.8±183.5 1600.1±222.8 1308.4±328.5
Pk to Pk Amp (uV) 73.1±16.5 44.83±16.9 48.3±10.8** 56.6±12.1
SpikeSegments 4735±2527 5823±3636 4598±3018 1206±1859**
Summary Table

24. Overall Conclusions
• Inter-electrode correlation (IEC) is noise resilient
• IEC is a pre-processing step: Can be used with threshold
detection and subsequent sorting methods
• May help reduce system requirements for downstream
operations (lower false positives)
• IEC may be useful in chronic neural interface
benchmarking studies because spike amplitude and
orientation are preserved for each electrode site

25. Future Considerations
• Correlation threshold for accept/reject can be
manipulated according to experimental need
• May incorporate spatial weighting matrix for
more closely spaced electrode sites
• Detect periods of high correlation and utilize
more intensive means to extract spikes

26. Acknowledgments
Penn State
• Timothy Gilmour (EE)
• Roger Gaumond
(emeritus)
• Jon Lawrence
External Consultants
• Steve Bierer (Univ.
Washington)
• Byron Olson
Partial Funding Provided By
• Penn State Department of Bioengineering
• National Institute of Deafness and
Communications Disorders (R21DC007227)

27. Thank you!
• Questions?

29. Receiver-Operator
Characteristic Curve Analysis
0 0.2 0.4 0.6 0.8 1
0.2
0.4
0.6
0.8
1
A
Low Common-Noise
0.2
0.4
0.6
0.8
1
A~0.82
Medium Common-Noise
0 0.2 0.4 0.6 0.8 1
0.2
0.4
0.6
0.8
1
B
~0.75
0 0.2 0.4 0.6 0.8 1
0.2
0.4
0.6
0.8
1
C
High Common-Noise
~0.82
False Positive Rate
TruePositiveRate
0 0.2 0.4 0.6 0.8 1
0.2
0.4
0.6
0.8
1
A
Low Common-Noise
0.2
0.4
0.6
0.8
1
A~0.82
Medium Common-Noise
0 0.2 0.4 0.6 0.8 1
0.2
0.4
0.6
0.8
1
B
~0.75
0 0.2 0.4 0.6 0.8 1
0.2
0.4
0.6
0.8
1
C
High Common-Noise
~0.82
0 0.2 0.4 0.6 0.8 1
0.2
0.4
0.6
0.8
1
A
Low Common-Noise
0.2
0.4
0.6
0.8
1
A~0.82
0 0.2 0.4 0.6 0.8 1
0.2
0.4
0.6
0.8
1
A
Low Common-Noise
0.2
0.4
0.6
0.8
1
A~0.82
Medium Common-Noise
0 0.2 0.4 0.6 0.8 1
0.2
0.4
0.6
0.8
1
B
~0.75
Medium Common-Noise
0 0.2 0.4 0.6 0.8 1
0.2
0.4
0.6
0.8
1
B
~0.75
0 0.2 0.4 0.6 0.8 1
0.2
0.4
0.6
0.8
1
B
~0.75
0 0.2 0.4 0.6 0.8 1
0.2
0.4
0.6
0.8
1
C
High Common-Noise
~0.82
0 0.2 0.4 0.6 0.8 1
0.2
0.4
0.6
0.8
1
C
High Common-Noise
~0.82
False Positive Rate
TruePositiveRate