The Partial Coherence method for assessment of impaired cerebral autoregulation using Near-Infrared Spectroscopy: potential and limitations

The Partial Coherence method for assessment of impaired cerebral autoregulation using Near-Infrared Spectroscopy: potential and limitations

S. Van Huffel*, J. Jacobs, D. De Smet*,J.

Vanderhaegen**, G. Naulaers** , M. Wolf***, P. Lemmers**** and F. van Bel****

KATHOLIEKE UNIVERSITEIT LEUVEN, BELGIUM

*DEPARTMENT OF ELECTRICAL ENGINEERING (ESAT-SCD)

** NEONATAL INTENSIVE CARE UNIT, UNIVERSITY HOSPITALS LEUVEN

*** NEONATAL INTENSIVE CARE UNIT, UNIVERSITY HOSPITAL ZURICH

****DEPT. OF NEONATOLOGY, UMC, WILHELMINA CHILDREN’S HOSPITAL UTRECHT

Introduction

Problem : impaired cerebrovascular autoregulation Δ CBF brain injuries

Preterm infants: high risk for development disorders

• Δ MABP frequent

• Δ MABP Δ CBF in some infants

To detect impaired autoregulation : Δ MABP Δ CBF

Acronyms :

• CBF : Cerebral Blood Flow

• MABP : Mean Arterial blood Pressure 1. Introduction

2. Methods 3. Models 4. Datasets 5. Results

6. Conclusions

Introduction

As rSO2, TOI, HbD reflect CBF under certain assumptions (ISOTT 07 & 08), study instead impaired autoregulation Δ MABP Δ HbD (TOI, rSO 2 ) with Δ SaO

= 0

Acronyms :

• HbD : cerebral intravascular oxygenation (=HbO

-HbR)

• SaO

: arterial oxygen saturation

• rSO 2 : regional cerebral oxygen saturation

Classical Method to quantify concordance: COH (Tsuji 2000) Aim :

• PCOH is proposed which essentially measures COH between modified signals (influence of SaO 2 removed)

• Compare different physiological models of SaO ² influence

• Compare results using three different datasets

1. Introduction 2. Methods 3. Models 4. Datasets 5. Results

6. Conclusions

1. Introduction 2. Methods 3. Models 4. Datasets 5. Results

6. Conclusions

PCOH

Start from x

(t)= SaO2, x

(t)= MABP, x

(t) =HbD (time domain) Step 1 : Fourier transform X

(f), X

(f), X

(f) (freq. domain) Step 2 : expression of the linear dependence between signals (black-box model)

With:

• P

: CSD between x

and x

• P

: PSD of x

( = ENERGY of x

)

• X

(f) : dependent on X

and X

• X

(f) : dependent on X

+ generalization to X

, X

, … and X

, X

, …

Step 3 : calculation of PCOH + generalization

Methods

Models

The classical method COH assumes the following model:

MABP versus HbD (TOI, rSO 2 )

We compared the following four models for PCOH:

1. Model 1: MABP \ SaO ² versus HbD \ SaO ² 2. Model 2: MABP versus HbD \ SaO ² 3. Model 3: MABP \ i(SaO ² ) versus HbD \ i(SaO ² ) 4. Model 4: MABP versus HbD \ i(SaO ² ) with i(SaO 2 ) = SaO 2 – f(MABP)

Remark: instead of HbD, TOI or rSO ² may be used

1. Introduction

2. Methods

3. Models

4. Datasets

5. Results

6. Conclusions

Datasets

30 preterm infants with need for intensive care monitored during first day(s) of life: MABP monitored by indwelling arterial catheter, SaO ² by pulse oximetry on limb.

• UZ Leuven (10): TOI (NIRO300, Hamamatsu), PMA=28.7±3.3 weeks, bodyweight=1125±504 g, recording time =

6h54min±3h33min

• UMC Utrecht (10): rSO ² (INVOS4100, Somanetics), PMA= 29.3

±1.3 weeks, bodyweight=1131±311g, recording time=49h36min±

11h48min

• UZ Zurich (10) : HbD (Critikon2001, GE), PMA=28.1±2.1 weeks, bodyweight=1198±439g, recording time=2h11min± 27min,

inspired O ² fraction changed SaO ² ↑

All data downsampled to 0.333Hz (3s) and (pre)processed as follows:

• If artifacts > 1.5 s, then exclude data else interpolate (Soul, 07)

• Data only included if SaO ² satisfies: 80%< SaO ² < 100%

• (P)COH restricted to 0.0033-0.04 Hz (25-300s cycles) (Soul,07)

• Sliding windowing, half overlapping epochs of 12.5, 15, 10 min

• Mean COH calibrated such that CSV=0.5 for impaired autoreg.

1. Introduction 2. Methods 3. Models 4. Datasets 5. Results

6. Conclusions

Results

Global Analysis

• On average, PCOH > COH (mean scores, CPRT* and PPI**), STD of PCOH values slightly higher but not significantly.

• Model 3 gives the highest values while model 2 the smallest ones for all datasets.

• High Mean score values (> 0.5) are related to smaller PMA and less significantly to smaller birth weight and smaller survival rate

• High PCOH values (mean>0.5, CPRT>0, PPI>0) better detect bad clinical outcome than COH (MDI<84, PDI<84, Apgar<7).

• CPRT and PPI (0.1) better detect bad clinical outcome than mean score values.

Local Analysis: zoom in on epochs with higher var(SaO 2 )

• All mean score values significantly higher, model 3 highest.

• Using unpreprocessed data, score values are all significantly lower (factor 2) and don’t reflect concordances between signals

*CPRT = % of Recording Time with mean (P)COH >0.5 (ISOTT 2007)

**PPI = % of 10 min. epochs with mean (P)COH > 0.68 (Soul et al, 2007)

1. Introduction

2. Methods

3. Models

4. Datasets

5. Results

6. Conclusions

COH model1 model2 model3 model4

mean ^{Leuven 0.39 0.41 0.41 0.44 0.42}

STD ^{Leuven 0.09 0.10 0.09 0.11 0.10}

CPRT ^{Leuven 14% 20% 14% 27% 22%}

PPI ^{Leuven 10.1% 12.4% 11.6% 18.3% 10.4%}

mean ^{Utrecht 0.35 0.34 0.33 0.39 0.36}

STD ^{Utrecht 0.09 0.09 0.09 0.10 0.09}

CPRT ^{Utrecht 13% 11% 8% 22% 17%}

PPI ^{Utrecht 20.8% 19.3% 15.1% 28.2% 21.5%}

mean ^{Zurich 0.56 0.61 0.57 0.63 0.59}

STD ^{Zurich 0.09 0.09 0.09 0.12 0.11}

CPRT ^{Zurich 72% 75% 70% 80% 71%}

PPI ^{Zurich 17.0% 34.0% 19.3% 40.1% 29.0%}

Local analysis COH model1 model2 model3 model4 ^{VarSaO 2}

Overall mean 0.55 0.52 0.52 0.61 0.57 1.20

Mean 20-40min 0.64 0.51 0.61 0.75 0.73 2.51

Local analysis COH model1 model2 model3 model4 ^{varSaO 2}

Overall mean 0.39 0.37 0.38 0.39 0.40 1.58

Mean 44-62min 0.42 0.51 0.49 0.55 0.50 3.05

Conclusions

•PCOH removes influence of SaO ² variations, thereby improving automated monitoring of (impaired) autoregulation .

•(P)COH analysis performed using three different datasets (different center, different CBF reflector)

• PCOH > COH: PCOH better detector of bad clinical outcome. CPRT and PPI (0.1) clearly better detectors than mean scores

•4 PCOH models compared  model 3 best

•Further statistical analysis needed, as well as correlation of new scores with neurological

outcomes on larger multicentric datasets

1. Introduction 2. Methods 3. Models 4. Datasets 5. Results

6. Conclusions