Predicting Respiratory Depression in Neonates Using Deep Learning

Neural Networks

Aleksandar Jeremic

and Dejan Nikolic

Department of Electrical and Computer Engineering, McMaster University, Hamilton, ON, Canada

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

depressions using physiological measure-

ments. We train deep neural network using previously obtained data set from NICU, McMaster University

Hospital with intra-arterial pressure measurements and evaluate its performance. Preliminary results indicate

that adequate performance can be achieved if sufﬁcient number of measurements is available.

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 fol-

low up studies of infants who received neonatal in-

tensive care show high rates of long-term respiratory

and neurodevelopment morbidity. As a consequence,

current research efforts are being focused on reﬁne-

ment 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

) and

carbon-dioxide (CO

) 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

partial pressure variations in the ar-

terial blood of a ventilated neonate. We evaluated the

applicability of the proposed model using 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 (Jeremic and Nikolic, 2019) we proposed

an algorithm for prediction of clinical depression

in neonates using parametric model based approach.

The proposed algorithms performs detection of pO

depression events using intra-arterial pressure mea-

surements and parametric model based on the log-

Riemannian distance between sample covariance ma-

trix measurements. However, intra-arterial pressure

measurements are administered only in rare cases that

warrant more intensive style of patient monitoring. To

this purpose in this paper we design a deep neural net-

work and use available data-set and intra-arterial pres-

sure measurements as ground truth in order to train

the network.

Deep neural networks are becoming increasingly

popular in the biomedical signal processing due to

the fact that they do not rely on the parametric model

which may be beneﬁcial due to patient-to-patient vari-

ability. In Section 2 we outline the structure of the

1054

Jeremic, A. and Nikolic, D.

Predicting Respiratory Depression in Neonates Using Deep Learning Neural Networks.

DOI: 10.5220/0013385400003911

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 18th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2025) - Volume 1, pages 1054-1057

ISBN: 978-989-758-731-3; ISSN: 2184-4305

0 500 1000 1500 2000 2500

Time[ms]

2 [kPa]

Figure 1: Sample of intra-arterial pressure measurements.

deep neural network and the corresponding signal

processing algorithms. In Section 3 we illustrate the

applicability of the proposed techniques by evaluat-

ing the performance of the proposed network using

a validation set that was not used as a part of train-

ing. Our preliminary results indicate that the onset of

oxygen depression can be predicted with potentially

useful accuracy sufﬁcient to alarm the clinical staff

of the NICU. In Section 4 we provide concluding re-

marks and discuss further directions.

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 McMaster University Hospital. The

data set consists of intra-arterial partial pressure mea-

surements obtained from 91 ventilated neonates. The

sampling time was set to 10s and the expiratory rate

was set to 1 breath per second. In Figure 1 we illus-

trate a sample of intra-arterial pressure measurements.

In order to predict the onset of respiratory de-

pression (hypo-ventilating) condition in (Jeremic and

Nikolic, 2019) we calculated the sample covariance

matrix we propose to use Frechet mean (Jahromi,

2014) which is given as the point which minimizes

the sum of the squared distances (Barbaresco, 2008):

and the log-Riemannian distance measure given by

(Moakher, 2005):

(A, B) =



log(A

−

)



∑

i=1

log

)

(1)

where the L

’s are the eigenvalues of the matrix A

−1

(Absil et al., 2009) where A and B are arbitrary M × M

matrices whose distance is being calculated. The de-

tails of this algorithm are provided in the aforemen-

tioned reference (Jeremic and Nikolic, 2019).

Figure 2: The schematic of the artiﬁcial neural network

(ANN).

Figure 3: Validation results for ANN training for different

scenarios.

In this paper we propose to use deep learn-

ing neural network (DNN) illustrated in Figure 2.

We use the four dimensional physiological measure-

ments consisting of partial pressure measurements

of oxygen, breathing rate, electrocardiography mea-

surements and intra-arterial pressure measurements

(available only for some patients). The covariance

matrix on the temporal window of N samples is cal-

culated of the physiological signals is calculated ( 4 x

4 ) and provided as a time-series input to DNN. The

training of the network is provided as an additional

input of the critical event (respiratory depression) is

labeled by a critical event triggered by an alarm which

is used as a ground truth. In Figure 3 we illustrate the

performance of the DNN to adequately predict the fu-

ture values of the distance measure (i.e. distance be-

tween sample covariance matrix of the physiological

signals). To illustrate the applicability of the proposed

method and evaluate the need for intra-arterial pres-

sure measurements we utilize two different designs

which differ with respect to input dimensions with

and without the presence of the intra-arterial pressure

measurements. The orange line represents validation

error in the absence of the intra-arterial pressure. The

purple, red and blue lines represent validation errors

for the DNN that are trained to predict critical event

in time windows of 1, 5 and 10 minutes respectively.

The above results indicate that in order to achieve

lower validation error we would beneﬁt greatly from

intra-arterial pressure measurements.

Predicting Respiratory Depression in Neonates Using Deep Learning Neural Networks

1055

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 91 patients with the num-

ber of recorded days for each patient varying between

1 and 90 with a mean of 32 days. In order to study

the changes with respect to the length of stay we se-

lected only patients who stayed up to 4 weeks and

performed evaluation for different weeks using cumu-

lative dataset.

In Figure 4 we illustrate the event corresponding

to the respiratory depression as well as the change in

the covariance matrix distance when calculated com-

pared to the Frechet mean in the absence of respi-

ratory depression. In our previous work (Jeremic

and Nikolic, 2019) we evaluated performance of the

model based approach whose application relies on the

use of one particular distance. We believe that in com-

plex physiological problems we may beneﬁt by allow-

ing the network to learn the processes in an uncon-

strained way as different distances may work better

for different patients. Since the time is of essence

in NICU, we may not have enough time to gather

enough data for a particular patients and thus deep

learning networks may be more useful regardless of

patient to patient variability. To this purpose the orig-

inal DNN was extended to include task of classiﬁca-

tion using the respiratory depression alarms as indi-

cated by cot-bed monitors.

In order to evaluate the performance of the pro-

posed DNN we train the network using the available

data for (N-1) i.e. 91 patients and evaluate perfor-

mance on the single patients that was not used for

training. We then repeat these process for all the 92

patients thus effectively performing rotational eval-

uation of the proposed network. In Figures 5 and

6 we illustrate the respiratory depression detection

for different subjects 5 min and 10 min before res-

piratory depression respectively. The four different

colours represent different weeks of NICU hospital-

ization. Preliminary results indicate that for this par-

ticular set there is not a signiﬁcant change in our abil-

ity to predict onset of respiratory depression. Please

note that this may be the consequence of the physio-

logical state of the patients which is general correlated

0 500 1000 1500 2000 2500

a) Time [ms]

2 [kPa]

0 500 1000 1500 2000 2500

b) Time[ms]

0.1

0.2

0.3

0.4

0.5

0.6

Distance

Figure 4: Sample of depression event as perceived by intra-

arterial pressure measurement and breathing rate.

Figure 5: Detection probability as a function of false posi-

tives 5min window.

with the length of stay. In Figure 7 we illustrate the

overall accuracy for all the subjects where the overall

accuracy is deﬁned as inverse of both types of error

(probabilities of false alarm and miss).

Finally in Table 1 we illustrate the precision, re-

call and F-score of the proposed system including all

91 patients. We can observe that the performance sig-

niﬁcantly deteriorates when compared to ”more sta-

ble” patients due to the fact that in the case of long

stay patients the number of respiratory depressions

is much larger. In addition, the statistical variability

of the measurements in these patients may be much

higher.

Figure 6: Detection probability as a function of false posi-

tives 10 min window.

BIOSIGNALS 2025 - 18th International Conference on Bio-inspired Systems and Signal Processing

1056

Figure 7: Overall probability of error per subject.

Table 1: Overall performance for all the patients.

Window Precision Recall F-score

10 min 0.7099 0.6911 0.7003

8 min 0.7492 0.733 0. 741

5 min 0.7846 0.7709 0.7777

3 min 0.7546 0.7411 0.7478

2 min 0.7874 0.7774 0.7824

1 min 0.8294 0.8201 0.8247

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. Due to the

fact that patient-to-patient variability and short length

of stay per patient may prevent utilization of paramet-

ric models we evaluate applicability of deep learning

networks. We illustrate the applicability of the pro-

posed method for patients with the length of stay of

up to four weeks. Our preliminary results indicate that

DNN could be potentially used in a hybrid setting in

which the rough estimate of model parameters could

be obtained using DNN and the ﬁne tuning for a par-

ticular patient could evolve throughout the patient’s

stay at NICU.

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.

Jeremic, A. and Nikolic, D. (2019). Predicting respira-

tory depression in neonates using intra-arterial pres-

sure measurements. BIOSIGNALS 2019, pages 342–

347–.

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.

Predicting Respiratory Depression in Neonates Using Deep Learning Neural Networks

1057