Comparison of coherence measures for assessment of impaired cerebral
autoregulation
D. De Smet*, J. Vanderhaegen**, G. Naulaers** and S. Van Huffel*
KATHOLIEKE UNIVERSITEIT LEUVEN, BELGIUM
*DEPARTMENT OF ELECTRICAL ENGINEERING (ESAT-SCD)
**NEONATAL INTENSIVE CARE UNIT, UNIVERSITY HOSPITALS LEUVEN
Introduction
Problem : defective cerebrovascular autoregulation
Δ CBF brain injuries
Premature infants : propensity for development because :
• Δ MABP frequent
• Δ MABP Δ CBF in some infants
1st mean to detect defective autoregulation :
Δ MABP Δ CBF
Acronyms :
• CBF : Cerebral Blood Flow
• MABP : Mean Arterial blood Pressure
1. Introduction
2. Datasets
3. Preprocessing
4. Sampling frequency
5. Welsh coherence
6. Critical score value
7. In the practice
8. Conclusion
Introduction
But, if Δ SaO2 = 0, then
Δ HbD Δ CBF(hypothesis)
2nd mean to detect defective autoregulation :
Δ MABP Δ HbD with Δ SaO2 = 0
Acronyms :
• HbD : cerebral intravascular oxygenation (=HbO2-HbR)
• SaO2 : arterial oxygen saturation
Aim : allow correction with medication such that Δ CBF=0
Method: The coherence coefficient is a measure of the linear dependence between two signals in the frequency domain.
1. Introduction
2. Datasets
3. Preprocessing
4. Sampling frequency
5. Welsh coherence
6. Critical score value
7. In the practice
8. Conclusion
Datasets
• More than 50 premature infants with need for intensive care from the hospitals of Zürich, Utrecht, and Leuven.
• MABP, SaO2, and the NIRS-measured HbD/rSO2/TOI measured simultaneously in the first days of life.
Acronyms :
MABP : mean arterial blood ressureHbD : cerebral intravascular oxygenatinonrSO2 : regional oxygen saturationTOI cerebral tissue oxygenationNIRS : near infrared spectroscopySaO2 : arterial oxygen saturation
1. Introduction
2. Datasets
3. Preprocessing
4. Sampling frequency
5. Welsh coherence
6. Critical score value
7. In the practice
8. Conclusion
Preprocessing
Each artifact point was deleted (Soul et al. 2007)
Step 1: keep signals within normal ranges
Preprocessing
Step 2: remove artifacts in MABP
Preprocessing
Step 3: remove artifacts in SaO2
Preprocessing
Step 4: remove artifacts in HbD
Preprocessing
Preprocessing makes the coherence growing (S. Van Huffel, iSOTT 08)
Preprocessing has bad consequence on the frequency content of the signals
Sampling frequency
Condition : sampling frequency (fs) > signal fluctuation frequency
Cyclical fluctuations in the case of continuously measured signals :
• CBV/HbTot : 2 to 4.7 cycles/min1 and 3 to 6 cycles/min2 (by NIRS)
• MABP : One cycle every 1 to 2.5 min1
fs > 0.1Hz
Acronyms :• CBV : cerebral blood volume• HbTot : total haemoglobin• MABP : mean arterial blood pressure_____________________________________________________[1] von Siebenthal et al., Brain & Development, 1999.[2] Urlesberger et al., Neuropediatrics, 1998
1. Introduction
2. Datasets
3. Preprocessing
4. Sampling frequency
5. Welsh coherence
6. Critical score value
7. In the practice
8. Conclusion
Step 1 : divide signal in epochs (C-windows). Divide each C-window in N segments (called H-windows as it is a highpass filtering)
Welsh coherence
Step 2 : (1) detrend, (2) apply Hanning windowing, and (3) compute the PSD/CSD for each H-window
Acronyms :
• Pxy(f) : crosspower spectral density (CSD) of x(t) and y(t) at a given frequency f
• Pxx(f), Pyy(f) : power spectral densities (PSD) of x(t), respectively y(t)
Step 3 : average the N modified H-windows
Step 4 : keep frequency band of interest (fCut : cutoff freq.)
Welsh coherence
Step 5 : compute average amplitude of spectrum in the frequency band of interest
REMARK 1 The FFT (fast Fourier transform) supposes the signals are periodical
REMARK 2 A complete period of the signal should be contained in a each H-window
REMARK 3The higher the value of N, the lower the
variance of the estimates of the spectra (SNR grows)
Problem : if N is too large, then the amplitude of the spectra diminishes1
N close to 82,3
_____________________________________________________[1] De Smet., Unpublished, 2007.[2] Kay, Prentice Hall, 1988 (book).[3] Taylor, Circulation, 1998.
Welsh coherence
1. Introduction
2. Datasets
3. Preprocessing
4. Sampling frequency
5. Welsh coherence
6. Critical score value
7. In the practice
8. Conclusion
REMARK 4All parameters satisfy
• TH acts as a highpass filer• The ratio TH/TC should be in the range of 0.51 or
smaller
REMARK 5The ratio THOver/TH should be equal to 0.5 if a
Hanning window was applied to the H-windows prior to the periodogram average2.
_____________________________________________________[1] De Smet., Unpublished, 2007.[2] Carter, IEEE Trans. on Audio and Electroacoustics, 1973.
Welsh coherence
1. Introduction
2. Datasets
3. Preprocessing
4. Sampling frequency
5. Welsh coherence
6. Critical score value
7. In the practice
8. Conclusion
Consequence : TC= TH(N+1)/2
TH and TH/TC are the sole parameters we really can choose
But … REMARK 6 TH have to satisfy :• TH < TC• TH > 10s1,2
Acronyms :• TC : duration of C-window (calculation window)• TH : duration of H-window (highpass filtering window)• N : number of averages in the Welsh method
_____________________________________________________[1] von Siebenthal et al., Brain & Development, 1999.[2] Urlesberger et al., Neuropediatrics, 1998
Welsh coherence
1. Introduction
2. Datasets
3. Preprocessing
4. Sampling frequency
5. Welsh coherence
6. Critical score value
7. In the practice
8. Conclusion
Critical Score Value
The critical score value (CSV) is the value above which the amplitude of coherence witnesses a significant linear concordance between the input signals.
Possible value for CSV are :• CSV=0.51
• or2
Acronyms : : significance level (e.g. 0.05) : to be chosen• d : 2.83*TC/TH (for Hanning window) : TH to be
chosen• F : F hypothesis testThe significance level greatly influences the CSV, in the case that the remainder parameters are unchanged!!_____________________________________________________[1] De Boer et al., Med. Biol. Eng. Comput., 1985.[2] Taylor et al., Circulation, 1998
1. Introduction
2. Datasets
3. Preprocessing
4. Sampling frequency
5. Welsh coherence
6. Critical score value
7. In the practice
8. Conclusion
Critical Score Value
Problem : this formula doesn’t take into account THOver (and thus N) that also has an influence on the amplitude of the coherence spectrum1
Solution : keep working with CSV=0.5 (two signals based on 50% shared variance),and calibrate2 mean COH (on all infants) on mean correlation coefficient (COR)
+ look at the range of COH and COR._____________________________________________________[1, 2] De Smet., Unpublished
1. Introduction
2. Datasets
3. Preprocessing
4. Sampling frequency
5. Welsh coherence
6. Critical score value
7. In the practice
8. Conclusion
Morren 1 Soul 2 Wong 3 Optimization
fs 6Hz (0.2Hz)
2Hz (O.4Hz)
1Hz >0.1Hz
(von Siebenthal,
Urlesberger)
TC 30min 10min 20min TH(N+1)/2
TCOver 10min 0min 0min E.g. : TC/2
N 217 3 5 close to 8 or calibr.
TH 12min 5min 10min > 2.5min
(von Siebenthal)
THOver 11min55s 2.5min 7.5min TH/2 if Hanning
(Carter)
fCut 0.01Hz 0.04Hz 0.02Hz <0.05Hz
(von Siebenthal,
Urlesberger, Nyq.)
CSV 0.5 0.77
(Taylor)
0.5 0.5 with calibr. or Taylor’s CSV
In the practice
[1] Morren et al., Proc. of the Intern. Conf. of IEEE, 2001.[2] Soul et al., Pediatric Research, 2007.[3] Wong et al., Pediatrics, to be published.
1. Introduction
2. Datasets
3. Preprocessing
4. Sampling frequency
5. Welsh coherence
6. Critical score value
7. In the practice
8. Conclusion
Conclusion
• nice preprocessing but bad consequence on the frequency content of the signals
• the problem of the varying amplitude of COH is solved by another manner Taylor did it. We showed Taylor’s method does not account for the overlap between H-windows
• we proposed optimized parameters to apply the coherence method
1. Introduction
2. Datasets
3. Preprocessing
4. Sampling frequency
5. Welsh coherence
6. Critical score value
7. In the practice
8. Conclusion
Thanks
toPhD grant Fin. ContributorsResearch Council KULeuvenFlemish GovernmentBelgian Federal Science Policy OfficeEUESA WorkgroupProf. Dr. Ir. S. Van HuffelProf Dr. G. NaulaersLic. J. Vanderhaegen