Identification of Patient Ventilator Asynchrony in Physiological Data

Through Integrating Machine-Learning

Anthony J. Stell

, Ernesto C. Caparo

, Zhe Wang

, Chenyang Wang

David J. Berlowitz

, Mark E. Howard

, Richard O. Sinnott

and Uwe Aickelin

School of Computing and Information Systems, University of Melbourne, 700 Swanston Street, Melbourne, Australia

Institute of Breathing and Sleep, Austin Hospital, 145 Studley Road, Heidelberg, Australia

Austin Health, 145 Studley Road, Heidelberg, Australia

chenyangwang3@student.unimelb.edu.au, {david.berlowitz, mark.howard}@austin.org.au

Keywords: Patient Ventilator Asynchrony, Machine-Learning, European Data Format (EDF).

Abstract: Patient Ventilator Asynchrony (PVA) occurs where a mechanical ventilator aiding a patient's breathing falls

out of synchronisation with their breathing pattern. This de-synchronisation may cause patient distress and

can lead to long-term negative clinical outcomes. Research into the causes and possible mitigations of PVA

is currently conducted by clinical domain experts using manual methods, such as parsing entire sleep

hypnograms visually, and identifying and tagging instances of PVA that they find. This process is very labour-

intensive and can be error prone. This project aims to make this analysis more efficient, by using machine-

learning approaches to automatically parse, classify, and suggest instances of PVA for ultimate confirmation

by domain experts. The solution has been developed based on a retrospective dataset of intervention and

control patients that were recruited to a non-invasive ventilation study. This achieves a specificity metric of

over 90%. This paper describes the process of integrating the output of the machine learning into the bedside

clinical monitoring system for production use in anticipation of a future clinical trial.

1 INTRODUCTION

Patient Ventilator Asynchrony (PVA) occurs where a

mechanical ventilator assisting a patient's breathing

falls out of synchronisation with their intrinsic

breathing pattern. This de-synchronisation may result

in patient discomfort and can lead to long-term

negative clinical outcomes. Three types of PVA that

have been demonstrated to impact on a patients

during non-invasive ventilation (NIV): 1) ineffective

effort - where the patient tries to take a breath, but this

effort fails to register with the ventilator, and it does

not provide the necessary support; 2) autocycle – a

small period of volatility where the patient takes

https://orcid.org/0000-0003-4819-9883

https://orcid.org/0009-0003-7499-0369

https://orcid.org/0000-0001-6054-6468

https://orcid.org/0000-0003-3217-1122

https://orcid.org/0000-0003-2543-8722

https://orcid.org/0000-0001-7772-1496

https://orcid.org/0000-0001-5998-222X

https://orcid.org/0000-0002-2679-2275

several breaths in quick succession and the ventilator

fails to respond to the rapid behaviour; 3) double

trigger - where the patient has taken two breaths, one

of which the ventilator fails to register (Hannan et al,

2019).

It has been shown that frequent PVA events

during both invasive- and non-invasive ventilation

can lead to many adverse consequences for a patient,

ranging from reduced sleep quality to more serious

outcomes such as lung injury, and an increased ICU

and hospital mortality rate (Brochard et al, 2014). It

is a significant burden on a variety of different patient

cohorts, such as those with specific conditions like

motor neuron disease (MND) or obesity

436

Stell, A., Caparo, E., Wang, Z., Wang, C., Berlowitz, D., Howard, M., Sinnott, R. and Aickelin, U.

Identiﬁcation of Patient Ventilator Asynchrony in Physiological Data Through Integrating Machine-Learning.

DOI: 10.5220/0012366700003657

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

In Proceedings of the 17th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2024) - Volume 2, pages 436-443

ISBN: 978-989-758-688-0; ISSN: 2184-4305

hypoventilation syndrome (OHS), but also severely

affecting those with more general chronic respiratory

failure. This group forms a significant percentage of

the global population.

Current practice to research and understand PVA

events includes clinical domain experts manually

assessing entire sleep hypnograms, along with

electroencephalogram (EEG) and physiological

output channels. For a single patient stay, this is often

of the order of several hours of data, with a resolution

of nanoseconds in some instances. The identification

of PVA events such as ineffective efforts, requires a

combination analysis of (for instance) the mask

pressure output (“pmask”) channel along with signs

of volatility in the EEG.

Technology used to tackle such problems at

present includes integrated tools, which present all

the information in one visual space that is convenient

and optimised for clinical use. An example of this is

the use of the CompuMedics sleep monitoring

software (Compumedics ProFusion, Abbotsford

Australia) that directly connects to the output from

the ventilator. However, despite the presence of this

integrated solution, given the relative frequency of

events against the time resolution described, this

process is highly labour-intensive and prone to error.

With recent advances in artificial intelligence and

machine-learning, it is clearly a process that would

benefit from automated optimisation. Therefore, this

project aims to integrate a machine-learning

algorithm that can automate the process of PVA

detection and provide clinical decision-support in the

form of suggestions of PVA event labels. These can

then be confirmed or rejected by the clinical domain

experts.

Several challenges in this work exist, which form

the basis of this paper. They include:

• integration of all data channels in an open

format supporting proprietary software such

as CompuMedics;

• reliance on hardware-accelerated processors

to fully exploit machine-learning algorithms;

• the need for specialist software libraries;

• the need to maximise memory efficiency when

choosing the software environment and

deciding on the system architecture;

• time-sampling factors such as down-sampling

of the machine learning output and the unit

choices of the EDF file specification, and

finally,

• the choice of presentation combined with the

operation of the algorithm to maximise the

utility for the clinical end-users.

2 BACKGROUND LITERATURE

Two interdisciplinary work threads have combined to

lead to the development of this work: one clinical and

the other from information and data science. The

clinical arm of this interdisciplinary group provided

the basis for this work through a randomised-

controlled trial that they had conducted previously,

where NIV was titrated with nocturnal

polysomnography (one of the first ever controlled

trials of this intervention) (Hannan, et al, 2019). One

of the primary findings of this study was that

polysomnography assisted optimization NIV titration

resulted in increased NIV usage (hours per night), and

an association was observed between fewer PVA

events and increased usage. In certain patient cohorts,

such as those living with motor neurone disease,

cohort evidence suggests that increased usage

(adherence to therapy) leads to a significant increase

in long-term survival (Berlowitz, et al, 2021).

2.1 PVA and NIV

Some independent software solutions to the issue of

detecting and mitigating PVA have been proposed

(Dres et al, 2021), but there is a general lack of

validation for these approaches and they still require

intensive effort to fully implement. Other studies

have focused on the storage of raw data on a long-

term basis (Janssens et al., 2015; Rabec et al., 2009)

but without estimations of signals and interpretation,

the utility of these tools have also yet to be

determined. This leads to an opportunity to explore

algorithmically-centered solutions integrated with

targeted software modules (as presented here).

Non-invasive ventilation (NIV) is a therapeutic

method used to provide respiratory support to

individuals with breathing difficulties without the

need for invasive procedures such as endotracheal

intubation. It delivers positive airway pressure to help

keep the airways open and assist with breathing. It

typically involves continuous monitoring to assess the

effectiveness and the patient’s response to treatment.

There are different techniques used for NIV

monitoring including capnography, which monitors

patient's exhaled carbon dioxide (CO2) levels.

A unique way of monitoring that captures the

interaction of patient and ventilator during nocturnal

use is polygraphy or polysomnography (PSG). It is

used in the context of sleep-related disorders, where

NIV can be titrated and monitored using

polysomnography. While ventilators normally only

provide readings such as mask pressure, PSG can

record various physiological parameters including

Identiﬁcation of Patient Ventilator Asynchrony in Physiological Data Through Integrating Machine-Learning

437

airflow, chest and abdominal movement (Gao, et al,

2021). Such data can help identify and classify PVA

events including Ineffective Effort (IE), Double

Trigger (DT) and Autocycle (AC).

A randomised controlled trial conducted by

(Hannan et al, 2019) suggested that using PSG to

titrate NIV therapy can lead to better alignment

between the patient's breathing patterns and the

ventilator's settings, but it may not reduce sleep

disruption. The data collected in this trial comprised

information from a cohort of 58 participants,

primarily individuals diagnosed with neuromuscular

disorders, all of whom were receiving NIV support

including the use of PSG titration.

2.2 Machine-Learning Approaches

A key unique aspect of the clinical work outlined in

this paper is that it has been performed during non-

invasive ventilation. This is a novel and ground-

breaking approach. Previous attempts at PVA

detection by other groups have always been

performed during invasive ventilation, usually in an

intensive care-unit. Therefore, when comparing

against other ML approaches to PVA detection, in the

invasive ventilation situation the available signals are

easier to detect as the system is closed, not open, and

thus inherently less noisy. It is in this context that

other ML approaches should be considered.

Zhang et al. (2020) proposed a novel method

using a two-layer neural network to detect the most

frequent types of PVA, resulting in the detection of

double triggering (DT) and ineffective inspiratory

effort (IIE). According to the study, it was shown that

ML-based approaches based on a robust database

(159 patients were included) could assist in PVA

recognition for clinicians.

Adams et al. (2017) explored the ventMAP

platform with focus on types of double-trigger and

breath stacking PVA. The algorithm proposed was

rule-based, using pressure and airflow signals,

including both a derivation and a validation cohort.

They obtained a performance of 92.2-97.7% on the

validation cohort. The algorithm helped in detecting

harmful forms of off-target ventilation in critical

patients.

The method developed by (Bakkes et al., 2020)

provided new insights for PVA. The study conducted

showed that the algorithm could detect and classify

types of PVA obtaining a precision average of 97.7%.

However, the study also emphasised the need for

inclusion of different network architectures to address

the necessary robustness of detection methods. It

should be noted that both algorithms (Adams et al.,

2017; Bakkes et al., 2020) faced different challenges

related to the data collection. Data labelling in the

(Bakkes et al, 2020) study was made by one expert

only, which led to an increase in the error margin of

the results. The platform ventMAP was capable of

obtaining a robust amount of data, however when

conducting the development and translation of the

output data into clinical applications, there were a

wide range of implementation issues (Adams et al.,

2017).

Another approach in the PVA field has been the

use of ensemble machine learning classifiers, e.g.,

(Rehm et al., 2018). The results suggest that high-

performing ML-based models are capable of

producing well-specified outputs despite the presence

of clinical artefacts. Therefore, the methodology used

serves as a helpful framework to guide classification

of such events.

3 METHODOLOGY

Considering the range of methods in the area, the use

of a data-driven approach has been embraced. (Wang

et al., 2022) proposed the use of several similarity and

randomness measures. This approach underpins this

paper, and specifically using variants of the matrix

profile (MP) algorithm. They achieved encouraging

results for detecting suspected PVA with a high

percentage test recall (90%+) among the reported

outputs. As a potential improvement on this

technique, the extension of similarity-based methods

to supervised nearest-neighbour search and including

techniques for ineffective effort detection has also

been considered here.

3.1 ML Algorithm

We extend Wang’s work (Wang et al, 2022) using an

algorithm that detects contiguous repeating patterns

in signals even with rhythm changes. This allows for

detection of abnormal changes, as well as

segmentation and feature analysis of the signals.

Follow-up models based on this algorithm have been

trained on an annotated non-invasive ventilation

waveform dataset, which gives a specificity and

sensitivity of over 90% in the context of detecting

auto-triggering events from noisy waveform data.

In the first instance, the practical measure of the

ML metric is simply the scalar number value

representing how volatile the different input channels

are (therefore, it is likely that a PVA event occurs near

a spike in the ML metric). However, a further future

refinement to this is to analyse the shape of the ML

HEALTHINF 2024 - 17th International Conference on Health Informatics

438

output and classify the type of anomaly based on that

shape.

The automated method to flag the anomalous

output is also based on several channels rather than

manually aligning and determining a PVA presence

based on the disjoint evaluation of each, which is

necessary when conducting the inspection manually.

The four primary channels used in the work are the

pressure mask (pmask), abdominal (abdo) and

thoracic (thor) respiratory inductance bands, and flow

as output form the NIV device (flow-tx). This multiple

channel feature is particularly useful in cases such as

ineffective efforts, which would not necessarily show

up on a single channel but could still be present.

3.2 Software Implementation

To integrate the output for the machine-learning

algorithm, an open-source format – the European

Data Format (EDF) specification (www.edfplus.info)

– has been chosen for data representation and

manipulation, not only because of the standardised

structure and well-supported open-source

community, but also to allow the portability of the

output across different platforms. This open format

supports the extraction in a standardised structure –

of both scored labels and free-text comments, which

have been added manually into the integrated bedside

system. It enforces a degree of structure on data

offering critical contextual data during a monitored

patient sleep. Such data is usually openly structured

and difficult to report in a standard way. Both scored

labels and free text annotations also present a

heterogeneity challenge in that there is no

standardized input approach. This effect is amplified

when there is more than one clinician involved in data

entry.

Therefore, the first step in producing the EDF file

with additional ML-output channel, is to extract the

original EDF file from the integrated bedside clinical

monitoring system, in this case CompuMedics, with

one EDF file per patient per stay. The meta-data

outlining the annotations accompanying that patient

stay are captured in the associated XML descriptor

file.

Once the original EDF has been extracted, the full

data integrity of that file is checked and written to

another newly created EDF file. This new file is

composed of the original physiological channels

chosen by the user with the addition of new channels

containing the ML-generated output.

This is achieved using the python library pyedflib,

which is a fork from the library EDFlib

(www.teuniz.net). These libraries are used to read the

EDF file’s properties and values including: number of

signals, channel indexes, sample frequency, and

number of data records. The physiological channels

are then read into this library for pre-processing, in

anticipation of processing by the ML algorithm. For

ease of persistent data storage, and to ease the burden

of volatile memory requirements, intermediate files

are written as part of these pre-processing steps.

These are stored in the Apache Parquet file format

(parquet.apache.org) - an open source, column-

oriented data file format, which uses in-built

compression for efficient data storage and retrieval.

The output of the parquet files is then fed into the

ML algorithm. The operation of the algorithm

involves many dependencies in both software and

hardware including:

• A memory-efficient version of the Anaconda

Python environment, known as MambaForge

(mamba.readthedocs.io), which provides a

setting that maximises the available underlying

memory to run the memory-intensive ML

algorithm.

• A library called Signatory that allows the

calculation of a “signature transform”, an

operation roughly analogous to a Fourier

transform that extracts information on the order

and area of a given data stream.

• Underlying GPU acceleration at a hardware

level, requiring the activation of an NVIDIA

processor (if available). In this project the Azure

cloud resource provides a VM within the “NC

series” that provides an NVIDIA GeForce RTX

3090 chip with 24 GiB

Using this software stack, once the algorithm is

fully computed, it is stored in a multi-dimensional

numpy array type, which the EDF file format and

libraries use heavily for functional operation. An

anomaly list is successively generated from the

numpy array, and this list is iterated over to produce a

sub-set of values where the ML metric has gone over

a user-supplied threshold number. This new subset

array is written to the same output EDF file but in the

form of point annotations.

In terms of timing, for each down-sampled

window, a corresponding value is produced

(measured in the arbitrary units of “ML-P”), along

with a corresponding timing value. This timing value

can be configured to be situated in the window at the

beginning (value 0), the end (value 1), the middle

(value 0.5), or any point in between. Due to the down-

sampling of the output by a factor of 16, the ML

output is rendered at a sample frequency of 2 Hz,

when combined with the pmask output in the final

Identiﬁcation of Patient Ventilator Asynchrony in Physiological Data Through Integrating Machine-Learning

439

EDF. This is due to the pmask output having a

frequency of 32 Hz (so when divided by 16, the final

frequency is 2 Hz). This output is written to the final

EDF using a pre-prepared buffer array of double

values, written to the file at repeat intervals governed

by that sample frequency. The sample frequency

itself varies according to channel and this is set at the

original point of first extraction of that channel from

the bedside system.

The EDF file itself is then rendered using open-

source tools that are freely available, again mainly

supported by the work of Teuniz van Beelen

(www.teuniz.net). The most popular option is

EDFBrowser, which provides standard tools to

operate and manipulate EDF files, such as varying

timescale, amplitude, window, and video

playback/recording. EDFBrowser is not entirely

portable across the main consumer operating

systems – for instance, installation on MacOS

requires detailed configuration that is a non-trivial

task for an average computer user. Therefore, an

alternative is the Polyman application

(sites.google.com/view/diegoalvarezestevez/projects

/polyman), which has similar tools but less visual

depth when rendering.

4 RESULTS

The presentation of the ML output was rendered

using EDFBrowser (figures 1 and 2).

Figure 1 shows the overall output when compared

between the machine-learning output (the green line)

and the entire output of a pmask channel for a 9.5 hour

patient stay (the yellow line, solid due to the high

resolution viewed from overall timepoint).

As can be seen, the highest points in the ML

output correspond to the largest variations in the

overall output of the pmask channel. In physical

terms, this most likely relates to the initial period

where the mask was not yet fitted to the patient, and

a period of adjustment that occurred midway through

the sleep.

From a clinical perspective, it is often the case that

the smaller, more subtle, variations in the ML output

are more useful. These are the “needle in a haystack”

points that the application is being used to identify,

rather than the large-scale volatility that can be most

easily seen on first viewing (though this helps to

validate that the ML output is in fact correctly

reflecting a valid physiological output). Therefore,

Figure 2 shows how the ML output looks when

compared against individual events in the

physiological channel, in this case one of the primary

PVA events: an ineffective effort.

Figure 2 also shows the annotations that are also

included in the output EDF file. These include a

combination of the threshold crossing notes (a

threshold of 4 was chosen for this exploration), as

well as the scored labels and free-text comments

extracted from the original EDF file. The high

number of these mean that the filtering tools available

through EDFBrowser needs to be used to

meaningfully navigate the EDF file and identify

points of interest in the readout (in EDFBrowser, the

list on the right of the window interactively

corresponds to the dashed line markers that span both

outputs).

5 DISCUSSION

There were a variety of issues that were encountered

when attempting to bring this implementation to a

production level. Many of those issues were due to

the interactions between uniquely-specified software

and hardware requirements that often led to

unpredictable interactions.

The Signatory library could only be run reliably

when executed on the Linux Ubuntu 22.04 operating

system. Portability across other operating systems

could not be guaranteed due, for instance, due to the

incompatibility with the Clang C-compiler, which

ships as standard on most MacOS versions.

Similarly, the Conda (Anaconda) Python

environment was required to set up and run the

dependency list to support the requirements for the

machine learning solution. However, due to the

extensive memory requirements, the larger scale

MambaForge environment was required. There are a

variety of flavours of this environment, which again

constrained the stack upon which the execution could

be performed.

The Pytorch library version – required to support

the complex functions of the Signatory library -

varied depending on whether the GPU or CPU

environment was available for execution. This

variation in itself created conflicting dependency and

version issues, which would need to be tightly

controlled before run-time, due to the need to

understand the particular hardware environment.

Overall, these issues could be grouped as the

requirement of pinned software versions along with

libraries without consolidated community support,

which is often a necessary feature of leading-edge

research.

HEALTHINF 2024 - 17th International Conference on Health Informatics

440

Figure 1: Overall monitored patient sleep, covering 9.5 hours. The large periods of volatility in the ML output (green)

correspond to aspects of the physiological output (yellow) which are highly differentiated from the majority of that output.

GPU requirements were also extensive, with a

minimum benchmark of 24 GiB in the “NC series” of

Azure VMs, required to execute the application on an

NVIDIA GeForce RTX 3090 processor. This was

required so that the calculation could be run over a

feasible timescale. There was a batch_size variable,

specifying the size of batches of data for processing,

which acted as an in-code handle and supported the

calculation accuracy and controlled the execution

time. But when this went over 10^6 (the minimum

requirement for sufficient accuracy), the calculation

time began to run to hours on a regular CPU platform.

However, again these are considerations that are

not uncommon in advanced ML approaches and will

inevitably become less problematic as processor

powers increase and execution speeds decrease. This

also has an impact on the cost-effectiveness of the

solution, e.g., does the cost of using such high-end

resources outweigh the cost of employing skilled

workers to manually detect PVA, and at what is the

trade off in accuracy? Such a cost-benefit exercise

would be a next logical step in evaluation of this

technology, repeated at various time intervals as the

underlying hardware improves.

Output timings were also a factor that require

further consideration. Due to the down-sampling of

output windows, some manipulation of the annotation

and ML output points was required, with a “stretching

factor” of 1.97 eventually settled upon. This was also

reflected in clinician feedback, where the annotation

indicating threshold crossing did not directly line up

with the corresponding point in the physiological

output, but rather it occupied a window, pre-

determined by the down-sampling rate. Though a

concern, it was noted that time was only one of

several factors that may have had an impact on the

system stability. Other factors such as the topography

of the ML output may influence the readout and allow

a classification of the type and presence of a given

instability. This could also help in identifying low-

resolution events, not immediately drawn out when

considering timing alone.

Finally, the idea of optimal presentation should be

considered. To express the output in the open-source

EDF format was a deliberate choice to promote

portability and accessibility. When taken to further

validation and downstream studies, the user-interface

considerations should be evaluated, and would

ultimately likely compete in terms of integration with

the in-situ bedside monitoring systems. If this

requires direct integration with the software vendor

solution this could become a block to further

development unless intellectual property and

collaborative agreements are negotiated.

Identiﬁcation of Patient Ventilator Asynchrony in Physiological Data Through Integrating Machine-Learning

441

Figure 2: This shows an example event of interest (an ineffective effort). The ML output corresponds to a higher-than-normal

spike in volatility, and the broken vertical line indicates that the annotation lining up in time with an effort that failed to

HEALTHINF 2024 - 17th International Conference on Health Informatics

442

6 CONCLUSIONS

In this paper, the implementation details of the

integration of a machine-learning algorithm to detect

PVA events, into a production version of a bedside

clinical environment has been presented. The

functional operation has been shown demonstrating

how the automated detection of ineffective efforts,

autocycles and double triggers can be achieved. We

also discuss the challenges encountered during the

work.

ACKNOWLEDGEMENTS

The authors would like to acknowledge the work of

Liam Hannan et al in providing the comprehensive

dataset upon which this work builds.

REFERENCES

Brochard L. et al, Measurement of esophageal pressure at

bedside: pros and cons. Curr Opin Crit Care 2014;

20:39–46

Martin Dres, Nuttapol Rittayamai b, Laurent Brochard

(2016). Monitoring patient–ventilator asynchrony

Candelaria de Haro, et al. (2018) Patient-ventilator

asynchronies during mechanical ventilation: current

knowledge and research priorities

Jean-Paul Janssens, Jean-Christian Bore, Jean-Louis

Pe´pin. (2015). Nocturnal monitoring of home non-

invasive ventilation: the contribution of simple tools

such as pulse oximetry, capnography, built-in ventilator

software and autonomic markers of sleep fragmentation

Rabec C, Georges M, Kabeya NK, et al. Evaluating

noninvasive ventilation using a monitoring system

coupled to a ventilator: a bench-to-bedside study. Eur

Respir J 2009;34:902e13

Lingwei Zhang et al. (2020) Detection of patient-ventilator

asynchrony from mechanical ventilation waveforms

using a two-layer long short-term memory neural

network.

Adams JY, Lieng MK, Kuhn BT, Rehm GB, Guo EC,

Taylor SL, Delplanque JP, Anderson NR (2017)

Development and validation of a multi-algorithm

analytic platform to detect off-target mechanical

ventilation. Sci Rep 7:14980

Bakkes, T. H., Montree, R. J., Mischi, M., Mojoli, F., &

Turco, S. (2020, July). A machine learning method for

automatic detection and classification of patient-

ventilator asynchrony. In 2020 42nd Annual

International Conference of the IEEE Engineering in

Medicine & Biology Society (EMBC) (pp. 150-153).

IEEE.

Wang, C., Aickelin, U., Luo, L., Ristanoski, G., Howard,

M. E., & Berlowitz, D. (2022). Patient-ventilator

asynchrony detection via similarity search methods.

Journal of Advances in Information Technology Vol,

13(1).

Rehm, G. B., Han, J., Kuhn, B. T., Delplanque, J. P.,

Anderson, N. R., Adams, J. Y., & Chuah, C. N. (2018).

Creation of a robust and generalizable machine learning

classifier for patient ventilator asynchrony. Methods of

information in medicine, 57(04), 208-219.

Berlowitz DJ, Sheers N. Not Only about the Drugs:

Improved Survival with Noninvasive Ventilation in

Amyotrophic Lateral Sclerosis. Ann Am Thorac Soc.

2021 Mar;18(3):419-420

Hannan LM, Rautela L, Berlowitz DJ, et al. Randomised

controlled trial of polysomnographic titration of

noninvasive ventilation. Eur Respir J. 2019 May

23;53(5):1802118

Xiang Gao, Yanru Li, Wen Xu, Demin Han, Diagnostic

accuracy of level IV portable sleep monitors versus

polysomnography for pediatric obstructive sleep apnea:

a systematic review and meta-analysis, Sleep Medicine,

Volume 87, 2021, Pages 127-137

Identiﬁcation of Patient Ventilator Asynchrony in Physiological Data Through Integrating Machine-Learning

443