Predicting Respiratory Depression in Neonates using Intra-arterial
Pressure Measurements
Aleksandar Jeremic
1
and Dejan Nikolic
2
1
Department of Electrical and Computer Engineering McMaster University, Hamilton, ON, Canada
2
Physical Medicine and Rehabilitation, University Children’s Hospital, Faculty of Medicine, University of Belgrade,
Belgrade, Serbia
Keywords:
Inter-arterial Pressure Measurements, Signal Models.
Abstract:
Respiratory problems are one of the most common reasons for neonatal intensive care unit (NICU) admis-
sion of newborns. It has been estimated that as much as 29% of late preterm infants develop high respiratory
morbidity. To this purpose invasive ventilation is often necessary for their treatment in NICU. These patients
usually have underdeveloped respiratory system with deficiencies such as small airway caliber, few collateral
airways, compliant chest wall, poor airway stability, and low functional residual capacity. Consequently ven-
tilation control has been subject of considerable research interest. In this paper we propose an algorithm for
detection of respiratory depression by predicting the onset of pO
2
depressions using intra-arterial pressure
measurements and second order statistical properties of these signals. We calculate the average covariance
matrix of intra-arterial pressure measurements in the absence of respiratory depression. We then use this
matrix as a reference measure and monitor the changes in the actual covariance matrix measurements. We
predict the onset of respiratory depression once the distance is larger than empirically determined threshold.
We demonstrate the applicability of our results using a real data set.
1 INTRODUCTION
Newborn intensive care is one of the great medical
success of the last 20 years. Current emphasis is
upon allowing infants to survive with the expectation
of normal life without handicap. Clinical data from
follow up studies of infants who received neonatal
intensive care show high rates of long-term respira-
tory and neurodevelopmental morbidity. As a conse-
quence, current research efforts are being focused on
refinement of ventilated respiratory support given to
infants during intensive care (Revow et al., 1989).
The main task of the ventilated support is to
maintain the concentration level of oxygen (O
2
) and
carbon-dioxide (CO
2
) in the blood within the phys-
iological range until the maturation of lungs occur.
Failure to meet this objective can lead to various
pathophysiological conditions. Therefore one of the
most critical components in the neonatal intensive
care units (NICU) is an adequate ventilation control.
In addition, due to a fragile state of neonatal lungs
the ventilation control has to be designed very care-
fully as neither hyperventilation nor hypoventilation
are acceptable.
In our previous work (Jeremic and Tan, 2007)
we developed a deterministic inverse mathematical
model of the CO
2
partial pressure variations in the ar-
terial blood of a ventilated neonate. We evaluated the
applicability of the proposed modelusing clinical data
sets obtained from neonatal multi-parameter intra-
arterial sensor which enables intra-arterial measure-
ments of partial pressures. Using this model we de-
veloped statistical signal processing model (Jeremic
and Tan, 2009) that predicts both inter-arterial pres-
sure measurements and corresponding confidence in-
tervals.
In this paper we present a new statistical signal
processing model for predicting the onset of respi-
ratory depression. We focus on advance detection
of falling edge events of oxygen partial pressure and
consequently rising edge events of carbon-dioxide
partial pressure measurements. To this purpose we
propose to detect the change by calculating distance
between the sample covariance matrix and reference
matrix obtained in the training stage. In Section 2
we describe the training procedure in which the ref-
erence estimate of the covariance matrix is calculated
and operating mode in which the actual sample co-
Jeremic, A. and Nikolic, D.
Predicting Respiratory Depression in Neonates using Intra-arterial Pressure Measurements.
DOI: 10.5220/0007576802370240
In Proceedings of the 12th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2019), pages 237-240
ISBN: 978-989-758-353-7
Copyright
c
2019 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
237
variance matrix is calculated using the sliding win-
dow approach. The sample covariance matrices are
calculated by using Frechet mean by utilizing our pre-
viously developed algorithms (Jahromi et al., 2015).
In Section 3 we present experimental results obtained
by applying our algorithms to real data set. Finally in
Section 4 we present conclusions and discuss possi-
bilities for future research.
2 SIGNAL PROCESSING
MODELS
To examine the applicability of the proposed algo-
rithms we apply them to the data set obtained in the
Neonatal Unit at St. James’s University Hospital. The
data set consists of intra-arterial partial pressure mea-
surements obtained from twenty ventilated neonates.
The sampling time was set to 10s and the expiratory
rate was set to 1 breathe per second. In Figure 1 we
illustrate a sample of intra-arterial pressure measure-
ments.
0 500 1000 1500 2000 2500
Time[ms]
5
6
7
8
9
10
11
12
p
O
2 [kPa]
Figure 1: Sample of intra-arterial pressure measurements.
In order to predict the onset of respiratory depres-
sion (hypo-ventilating) condition we propose to cal-
culate the sample covariance matrix and compare it to
a reference value obtained from the training period.
In order to achieve this goal we first select a training
data set by selecting time intervals without respira-
tory depression event. Note that these intervals have
to be selected for every patient by a trained neonatolo-
gist. Then we calculate the sample covariance matrix
by using the sliding window over the measured intra-
arterial pressure measurements. To determine the size
of the window we first calculate the autocorrelation
function, see Figure 3.
Note that there are various techniques for detect-
ing signal jumps / edges in arbitrary signals. However
since our goal is to predict the onset of these events,
i.e. detect conditions that lead to hypoventilation we
propose to use second-order statistical properties in-
1 2 3 4 5 6 7 8 9 10
Lag
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Autocorrelation
Figure 2: Autocorrelation function.
stead of simple trend changes. Namely, our prelimi-
nary results indicate that trend changes are frequent in
intra-arterial pressure measurements due to the nature
gas exchange and hence lead to significant number of
false positives.
In order to calculate the sample covariance ma-
trix we propose to use Frechet mean (Jahromi, 2014)
which is given as the point which minimizes the sum
of the squared distances (Barbaresco, 2008):
ˆ
S = argmin
S M
n
i=1
d
2
(S
i
,S ) (1)
where {S
i
}
n
i=1
represents the symmetric positive defi-
nite matrices and d(., .) denotes the metric being used
respectively. Therefore the above expression can be
interpreted as a way of calculating an averaged sam-
ple covariance matrix using a sliding window where
S
i
represents an ith window sample covariance es-
timate.
In this paper we propose to use log-Riemannian
distance measure given by (Moakher, 2005):
d
l
(A,B) =
log(A
1
2
BA
1
2
)
2
=
s
M
i=1
log
2
(L
i
)
(2)
where the L
i
s are the eigenvalues of the matrix A
1
B
(Absil et al., 2009) where A and B are arbitrary M× M
matrices. In our particular application matrix A refers
to sample covariance matrix of the measured signal
and B is the reference covariance matrix obtained in
the training phase i.e. on an interval where no depres-
sion events were recorded.
In order to determine the threshold for distance
comparison between the reference and testing co-
variance matrices we propose to use multiple signal
segments and calculate multiple reference covariance
matrices. Using these matrices we then calculate the
histogram of mutual distances and pick the thresh-
old based on the desired percentage of false positives.
Note that as a preliminary approach we choose to use
BIOSIGNALS 2019 - 12th International Conference on Bio-inspired Systems and Signal Processing
238
the same distance measure both for training and eval-
uating performance although other distance measures
(e.g. square root, Frobenius norm) could be used as
well.
Let y
i
be the i-th sample of inter-arterial pressure
measurements. Then the outline of the algorithm is as
follows
within the training data set create windows
~
data
k
= [y
(k1)l1+1
,· ·· , y
kl11
] where l1 is the
length of the window
within the above window select subwindows of
length l2 and label them
~
data
j
k
where j = 1, l1
l2+ 1
remove the sample mean from the window vectors
calculate rank 1 sample covariances
~
data
j,T
k
~
data
j
k
and average them using Frechet mean instead of
commonly used addition
3 RESULTS
We evaluate the performance of the proposed algo-
rithms using the data set obtained at the Neonatal
Intensive Care Unit, at McMaster University Hospi-
tal. At each cot in the Neonatal Intensive Care Unit
at the McMaster Children’s Hospital there is a cot-
side monitor displaying the physiological parameters
measured. Moreover, these monitors are linked in a
network with a central station into which data can be
rapidly exported via Draeger Infinity Gateway soft-
ware. Data is stored in the central station for 24 hours.
The final dataset contained 16 patients with the num-
ber of recorded days for each patient varying between
1 and 90 with a mean of 32 days.
In Figure 4 (a) we illustrate the event correspond-
ing to the respiratory depression and in Figure 4 (b)
we illustrate the corresponding change in the covari-
ance distance calculated using the aforementioned
logarithmic Riemannian distance. In Figure 5 we il-
lustrate the same results but in the absence of respira-
tory depression. Note that the dynamic range of the
distance is much smaller compared to Figure 4.
As we stated before the performance of the algo-
rithm will be dependent on adequate threshold selec-
tion. As a preliminary approach we arbitrarily set the
false positives rate to 10% and calculated the corre-
sponding threshold empirically using the part of data
set (for each patient) without hypoventilation events.
Using sliding window we calculate multiple sample
covariance matrix estimates. Then we use the testing
part of the data set and detect onsets of hypoventila-
tion events by comparing distance between training
0 500 1000 1500 2000 2500
a) Time [ms]
4
6
8
10
12
p
O
2 [kPa]
0 500 1000 1500 2000 2500
b) Time[ms]
0
0.1
0.2
0.3
0.4
0.5
0.6
Distance
Figure 3: a) Sample of inter-arterial partial pressure b) Co-
variance distance calculated using logarithmic measure.
200 400 600 800 1000 1200 1400
a) Time[ms]
4
6
8
10
12
p
O
2 [kPa]
0 500 1000 1500
b) Time[ms]
0
0.002
0.004
0.006
0.008
0.01
0.012
Distance
Figure 4: a) Sample of inter-arterial partial pressure b) Co-
variance distance calculated using logarithmic measure.
set covariance estimates to reference covariance ma-
trices. In both cases we calculate sample covariances
using sliding window of 100 samples and within that
window we use 5 samples to construct 5-dimensional
vector of lagged measurements which are then aver-
aged using Frechet mean in order to calculate sample
covariances.
In order to evaluate the performance of the pro-
posed algorithm we count the number of detected on-
sets and compare it to the actual count performed by
trained physicians. In Figure 6 we illustrate the per-
formance of our algorithm as a function of the false
alarm positives.
0 2 4 6 8 10 12 14 16 18 20
False positives [%]
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Detection probability
Figure 5: Detection probability as a function of false posi-
tives.
Predicting Respiratory Depression in Neonates using Intra-arterial Pressure Measurements
239
In Figure 6 we illustrate the performance mea-
sures for all the patients. Note that due to the
large patient-to-patient variability the performance
may vary significantly. This may partially be due
to the fact that the respiratory patterns are very de-
pendent on the gestation length and hence an effort
should be made in selecting adequate training sets.
0.06 0.07 0.08 0.09 0.1 0.11 0.12 0.13 0.14 0.15
Pf
0.74
0.76
0.78
0.8
0.82
0.84
0.86
0.88
0.9
0.92
Pd
Figure 6: Scatter plot of performance parameters for all the
patients.
4 CONCLUSIONS
One of the most important tasks that affect both long-
and short-term outcomes of neonatal intensive care is
maintaining proper ventilation support. To this pur-
pose in this paper we develop signal processing algo-
rithms for predicting the onset of hypoventilation in
order to increase efficient control of ventilation sys-
tem in timely manner. This is especially important for
neonates due to a fragile state of their lungs and hence
predicting the decrease oxygen levels can potentially
enable us to control the ventilator with smaller dy-
namic range.
In this paper we propose to predict the onset us-
ing second order statistical properties by calculating
sample covariance matrices using Frechet mean. Our
experimental results indicate that the structure of co-
variance matrix is slowly changing once the hypoven-
tilation begins. Due to the fact that the trend changes
of intra-arterial pressure occur continuously they may
not serve as a good indicator due to a large number of
false positives. To this purpose we focus our attention
on the second order properties i.e. covariance matrix
and utilize Frechet mean as it is know to be able to
capture different information about matrix structure
depending on the distance measure used. We evalu-
ate the performance of our algorithm using a real data
set previously labeled by trained physicians. In fu-
ture work we propose to develop multichannel infor-
mation fusion system that will use different distance
measures as onset detectors. In addition, we will com-
pare performance of our algorithm versus threshold-
ing. Note however that currently used thresholding
algorithms detect depression once it starts to occur
while our algorithm attempts to predict the onset of
the depression.
REFERENCES
Absil, P.-A., Mahony, R., and Sepulchre, R. (2009). Opti-
mization algorithms on matrix manifolds. Princeton
University Press.
Barbaresco, F. (2008). Innovative tools for radar signal pro-
cessing based on cartan’s geometry of spd matrices &
information geometry. In Radar Conference, 2008.
RADAR’08. IEEE, pages 1–6. IEEE.
Jahromi, M. (2014). Frechet means with respect to rieman-
nian distances: Evaluations and applications. Ma.Sc.
Thesis.
Jahromi, M., Jeremic, A., and Wong, K. (2015). Estimating
positive definite matrices using frechet mean. Biosig-
nals, 3:295–299.
Jeremic, A. and Tan, K. (2007). Estimating respiratory pa-
rameters using intra-arterial partial pressure measure-
ments. Asilomar, 2:851–854.
Jeremic, A. and Tan, K. (2009). Estimating respiratory
parameters using intra-arterial partial pressure mea-
surements and stochastic differential equations. 2009
Canadian Conference on Electrical and Computer
Engineering, 2:559–564.
Moakher, M. (2005). A differential geometric approach
to the geometric mean of symmetric positive-definite
matrices. SIAM Journal on Matrix Analysis and Ap-
plications, 26(3):735–747.
Revow, M., England, S., Beirne, H., and Bryan, A. (1989).
A model of the maturation of respiratory control in
the newborn infant. IEEE Trans. on Biomedical Engi-
neering, 36(4):415–423.
BIOSIGNALS 2019 - 12th International Conference on Bio-inspired Systems and Signal Processing
240