Predictive Model of Septic Shock Staging Base on Continuing

Invasive Hemodynamic Monitoring

Ruowen Liao

Department of Medical Biophysics, Western University, London, Ontario, Canada

Keywords: Septic Shock, Hemodynamic, Deep Active Learning, Multi-Classification, Predictive Model.

Abstract: Septic shock is a major public health concern across the world, also is a typical cause for patients being

admitted to the intensive care unit. It is easier to be misdiagnosed, yet the situation is getting worse. Septic

shock can be classified into three stages: irreversible (early stage), compensated, and decompensated. Sepsis

has long been misdiagnosed, but it develops and worsens at an alarming rate, often reaching irreversible

levels within hours. This work has expanded the proportion of invasive hemodynamics to septic shock for

the development of understanding of the phases of septic shock. This article aims to construct and develop a

real-time prediction model of septic shock staging based on continuous invasive hemodynamic monitoring.

The ultimate model of the article is a multi-classification prediction model.

In this experiment, the eICU collaborative research database was employed, and four characteristics from

the dataset were scored to indicate the stage of septic shock. Need to point out that deep active learning, a

new approach that combines deep and active learning, was chosen as the research's major learning approach.

Margin sampling is the main query strategy used in the active learning approach, with the random selection

strategy serving as a control strategy. There are two groups of query strategies, compare the two groups to

see which one is more effective: random selection or active learning. As a result, the query strategy of active

learning is considerably most stable than random selection in deep active learning. Although septic shock

cannot be diagnosed purely based on hemodynamic characteristics, the model can nevertheless assist

clinicians in making an early diagnosis or warning.

1 INTRODUCTION

Septic shock is a common reason for patients to be

admitted to the intensive care unit (ICU), and it is

also a significant cause of mortality among severely

sick patients in the ICU. In 2017, there were 48.9

million instances of sepsis and 11 million fatalities

due to sepsis, accounting for roughly 20% of all

deaths worldwide (Genga 2017). It is worth noting

that COVID-19, from its emergence in 2019 and

continues to this day, has been linked to sepsis.

Health care personnel pay particular attention to the

development of sepsis after a COVID-19 patient is

brought to the ICU (Bediako 2021). This was

demonstrated in many studies that shock could be

divided into three stages: irreversible, compensated,

and decompensated shock. Sepsis is very easy to be

misdiagnosed, but it deteriorates very quickly in

hours. The staging of shock assists medical

personnel in determining the severity of the

condition and appropriately intervening in treatment

and medicines to enhance patient survival rates.

There is thereby a need for classification, but it is

still a significant challenge to define clearly what

stage of shock the patient is at based on the clinical

presentation. Thus, to have better knowledge of the

phases of septic shock, this research has increased

the proportion of invasive hemodynamics in septic

shock.

The ultimate objective of this study is to provide

some help in using a machine learning approach to

determine the stage of shock in patients with sepsis

in the ICU, improve efficiency and reduce fatality.

There are many excellent reviews in literature

dealing with the basic concepts of machine learning

and sepsis. Continuing to learn about septic shock

using machine learning is also a significant step

forward in medicine. Notably, hemodynamic

monitoring is critical for the diagnosis and

intervention of septic shock patients. The eICU

Collaborative Research Database (eICU-CRD)

Demo was utilized as a source of clinical study data

518

Liao, R.

Predictive Model of Septic Shock Staging Base on Continuing Invasive Hemodynamic Monitoring.

DOI: 10.5220/0011373200003438

In Proceedings of the 1st Inter national Conference on Health Big Data and Intelligent Healthcare (ICHIH 2022), pages 518-523

ISBN: 978-989-758-596-8

in this study, which comprised 24 hours of

continuous vital sign monitoring of systemic

circulation (Badawi 2018). The author recruited

patients previously diagnosed with sepsis from the

eICU-CRD Demo for observation and research. The

use of machine learning for invasive continuous

hemodynamic monitoring of eICU sepsis patients is

expected to further improve the understanding of the

shock stage.

2 METHODS

2.1 The eICU-CRD Dataset

Between 2014 and 2015, researchers from the MIT

Computational Physiology Laboratory, Philips

Healthcare, and PhysioNet's colleagues collected

data from over 200,000 ICU patients for the ICU-

CRD database (Badawi 2018). It should be pointed

out that this database is an electronic version that

provides a new model of care in ICU: remote

monitoring. The e-recording allows clinicians to

instantly retrieve a patient's vital signs, saving time

and preventing the loss of paper data. This research

utilized the eICU-CRD demo as the experimental

database because the researcher intends to see if it

can generate predictions with a smaller amount of

electronic data. The eICU-CRD demo includes

2,500 patients in the ICU department from 20 large

hospitals in the United States. These patients are

divided into a training set and test set according to

the radio of 8:2. The file in the eICU-CRD called

'vitalPeriodic.csv' is particularly attractive as the

main dataset, due to the study is based on the

characteristic of hemodynamic to make a prediction

model. The VitalPeriodic table includes the

continuous invasive hemodynamic monitoring

features which are need in this research: heart rate,

oxygen saturation (SaO

), central venous pressure

(CVP), systolic blood pressure, and diastolic blood

pressure.

2.2 Features and Score Setting

Four features were collected to determine the stage

of septic shock: heart rate, CVP, mean arterial

pressure (MAP), SaO2.

2.2.1 Heart Rate

When cardiovascular decomposition occurs, the

heart is the first compensation mechanism. At this

time, the heart rate will increase to ensure sufficient

cardiac output. According to the definition and

diagnostic criteria of sepsis and septic shock, a heart

rate of more than 90 beats per minute or two

standard deviations greater than the normal value of

the same age can be confirmed or suspected of

infection(CCM1993).

2.2.2 CVP

It is generally believed that CVP at 8 to 12mmHg is

a treatment target for severe infections and septic

shock. In recent years, CVP has been challenged as

a pressure indicator to evaluate volume load. It is

now believed that CVP can be used to determine the

type of shock. However, unless in the extreme range

of the variables, such as in the case of a history of

bleeding, and the CVP value is 0mmHg, it should

always be interpreted together with other variables

(Antonelli 2014).

2.2.3 MAP

Invasive blood pressure (IBP) is a commonly used

technique in the ICU. Continuous monitoring as one

of the advantages of IBP could provide patients

status in real-time. In our research, MAP is selected

as a variable shown the IBP’s feedback of patients.

2.2.4 SaO2

As an important monitoring indicator of severe

infection and septic shock recovery, SaO2, also

selected as one of the scoring indicators in this

article. SaO2 value is from 60% to 80% in patients

with severe infection and septic shock in normal

circumstances. It must also be mentioned that a

significant increase in mortality when the SaO2

value is less than 70%.

2.2.5 Scoring Design

The designer created a simplified score sheet based

on the given information and the MAP data in the

APACHE II score, as shown in table 1.

Table 1: Criteria for scores calculated based on invasive

hemodynamic data patients.

Predictive Model of Septic Shock Staging Base on Continuing Invasive Hemodynamic Monitoring

519

Parameters Points

+4 +3 +2 +1

Heart Rate

(BPM)

- - - ≥90

CVP

(mmHg)

- - -

<8 &

>12

MAP

(mmHg)

≥160 or

≤49

130~159

≥110 or

≤ 69

SaO2 (%) 60~70 70~80 - -

In addition, the scores are divided as follows

based on the aforementioned features and scores to

identify the phases of septic shock: 1) a score of 0 to

4 is judged to be a non-septic patient. The patient’s

septic shock phase is assessed to be more severe as

the score rises. 2) with the score of 4, it is in

irreversible stage, 3) it is belonging to a

compensated stage when the score is 5 to 7, 4) and

the patient will be classified as in the

decompensated phase with the score of 8 to 10. This

shown as below figure. This is the multi-

classification standard of this experiment.

Table 2: Stratification criteria for multi-classification scores.

Score

Non-sepsis

Irreversible

Compensated

Decompensated

2.3 Model Development

The eICU-CRD is collected patients’physiological

data every few minutes. As mentioned previously,

the patients are randomly separated into two parts:

80% for model training (2000 patients), 20% as the

test set. Convolution neural networks (CNN) also is

a multi-layer neural network, were utilized in this

research to create a prediction model of which phase

of the septic shock the patient will be in. CNN has

the ability to extract features automatically, the

convolution layer is in charge of extracting features

and convolution is used to extract the needed

information (Asafuddoula 2016).

One of our goals is to develop a prediction

model with as minimal data as feasible; the model

also uses deep active learning (DAL), a hybrid of

deep and active learning(Chang 2020). The flow of

the model is shown in Figure 1 below. DAL

framework can be roughly divided into two parts:

the active learning query strategy on the unlabeled

data set and the training method of the deep learning

model(Chang 2020). There are hundreds, if not

thousands, of records for each of the 2,500 patients.

Unlike most traditional active learning algorithms,

which query one by one, batch model deep active

learning (BMDAL) picks an entire batch of

unlabeled data based on certain selection criteria

(Agarwal 2019). The amount of information and

diversity of the samples are considered at the same

time as the batch selection of samples (Agarwal

2019). The DAL code is built on a Github library

called "deep active learning" and is publicly

available. It is worth pointing is that the optimizer of

the DAL used Adam. The benefit of DAL code is

that it has a rapid gradient to huge data, which is

ideal for our needs in eICU-CRD, where we need to

analyze enormous volumes of data. Another

significant advantage of the DAL is that it does not

boost the budget of recognition and classification

(Chang 2014).

ICHIH 2022 - International Conference on Health Big Data and Intelligent Healthcare

520

Figure1: The framework of the deep active learning

process.

2.4 Query Strategy

The most important things in active learning are

how to select samples for labeling and the selection

of query strategy. There are two main principles for

how to select samples for labeling: uncertainty

principle and difference principle. Margin sampling

as one of the uncertainty samplings was chosen in

the active learning. The concept of active labeling

by margin sampling is to give the sample the

smallest separation between the top two class

predictions, as seen in the equation:











 (y















)

(1)

where y 1 and y 2 are the deep learning network's

first and second most probable class labels,

respectively (Agarwal 2019). The Random sampling

strategy, is mainly as a control strategy. This

strategy randomly select a certain proportion of

samples from the unlabeled samples and submit

them to the labeler for labeling. It also has been used

in this article. In the processing, there are two

groups:

One uses a combination of margin sampling and

the Random sampling, referred to as active learning

by learning strategy. Another does not use a margin

sampling strategy.

To determine which strategy is more successful,

the two groups will be compared in the next and the

more effective strategy will be found between

random selection and active learning.

3 RESULTS AND ANALYSIS

According to information derived from the eICU-

CRD sub-database ‘diagnose’, the proportion of

people diagnosed with septic shock during their ICU

stay was 7.33 percent of the total number of people

in the database. The percentage of positive to

negative occurrences was 3:38. These are patients

who were diagnosed with sepsis and shock by

clinicians, but who were not classified as being in

the septic shock phases by the doctors.

After 15 rounds with 1000 batch sizes and 1000

queries per round, the ultimate accuracy of the

active learning by learning approach is 55.017

percentage, whereas the accuracy of random

selection is 51.958 percent. Regardless of the fact

that there may not be much of a difference in the

accuracy, Figure 2 illustrates that the accuracy of

random selection outperforms that of active learning

initially. As shown in the diagram, the initial

random selection approach has significantly higher

accuracy and stability than the active learning

technique. However, when additional input data and

data are labeled, the stability of random selection

tends to deteriorate, compared to active learning.

Figure 2: Active accuracy of two query strategies.

Conversely, since more data is added to the

model, the accuracy of active learning gradually

overcomes the measurement rate of random

selection, and the active learning model is more

stable than random selection. This article also makes

use of deep learning. Deep neural networks (DNN)

were constructed, and the batch-based sample query

approach was used. The following graphs of

accuracy and loss are obtained using the case of a

batch size of 64. Deep learning is better than active

learning if readers look objectively at the accuracy

and loss of the final model in this experiment, as is

demonstrated in figures 3 and 4.

Predictive Model of Septic Shock Staging Base on Continuing Invasive Hemodynamic Monitoring

521

Figure 3. Deep learning accuracy.

Figure 4. Deep learning loss.

4 DISCUSSION

The most direct charm of active learning is that it

can significantly reduce the cost of labeling samples.

Researchers have discovered that accuracy is quite

low and about 0.55 based on of Figure 2, which

represents accuracy in both query techniques,

random selection and active learning. for this, the

authors have come up with the following hypothesis.

Firstly, it can no longer learn from the data, although

it is an active learning model. Our deep learning

model, on the other hand, appears to contradict this

notion. Furthermore, the authors discovered that the

percentage of individuals with septic shock in the

eICU-CRD demo is extremely tiny. The percentage

of scores below 4 that indicate a non-sepsis state is

approximately 0.99, and the distribution is

extremely unbalanced. As a result, it's thought that

this data set isn't appropriate for the test model, and

the accuracy rate is poor. Even though the testing

accuracy is indeed very high for this type of

database, it really has no effect on the model. A

further option is that a deep active learning library

setting was not properly debugged in the

experiment, preventing the DAL model from

learning anything.

Another thing worth mentioning is that because

of the interdependence of the sympathetic and

parasympathetic nervous systems, shock should not

be judged simply based on "normal" hemodynamic

measurements. Regardless of the fact that the model

can predict a patient's sepsis stage based on current

eICU datasets, it still has to be validated and

adjusted before it can be utilized in real life. Since

this complex septic shock phase is solely assessed

by invasive hemodynamics, the model is still

immature. Nevertheless, most ICU patients who

underwent intubation medication are adept at

gathering invasive hemodynamics parameters. As a

consequence, the model can still be used as a guide

to assist clinicians in promptly diagnosing or

warning of the onset of more severe shock.

5 CONCLUSIONS

Based on hemodynamic data and the features of

most intubation treatments in ICU patients, this

paper provides a technique for predicting and

staging septic shock in the article. For multi-class

prediction, deep active learning and deep learning

active learning are used to study. Deep learning

validates the model’s feasibility and correctness.

The query strategy of active learning is considerably

most stable than random selection in deep active

learning. The fraction of patients with sepsis is too

small since the data is concentrated, resulting in the

low accuracy of the active learning model. The low

accuracy and instability of the DAL model are

caused.

However, this paper also has the deficiency that

the author's knowledge of the DAL source code is

incomplete and inaccurate, a representative database

should be chosen to debug the model and code. And

even though it cannot fully diagnose and forecast

septic shock with invasive and continuous

hemodynamic monitoring of patients, this

experiment is likely to increase the understanding of

the shock stage and aid clinicians in quick diagnosis

and real-time prediction. Even if invasive

hemodynamics cannot properly detect and

discriminate the stages of septic shock after

successful debugging of the future model, it will

ICHIH 2022 - International Conference on Health Big Data and Intelligent Healthcare

522

offer a new research avenue for the study of the

stages of septic shock. Septic shock may be

immediately interfered with and the mortality

incidence of septic shock can be reduced by

accurately evaluating the stage of septic shock and

offering assistance to medical personnel.

ACKNOWLEDGMENTS

Professor Robert F. Murphy of Carnegie Mellon

University and Assistant Teacher Jinzhe Zhang of

the University of Tokyo are gratefully received for

their support and suggestions. Thanks for providing

the PhysioNet eICU-CRD data and the deep active

learning source code supplier. Simultaneously, I

express gratitude to the Western University for its

nurturing and the content related to this article was

learned in the department of medical biophysics.

Appreciate also the Editor and reviewers for your

valuable suggestions, which helped the author

enhances the work.

REFERENCES

Agarwal, A., Ash, J. T., Krishnamurthy, A., Langford, J.,

& Zhang, C, (2019). Deep batch active learning by

diverse, uncertain gradient lower bounds. arXiv

preprint arXiv:1906.03671.

Antonelli, M., Beale, R., Bakker, J., Cecconi, M., De

Backer, D., Hofer, C., Jaeschke, R., Mebazaa, A.,

Pinsky, M. R., Rhodes, A., Teboul, J. L., Vincent, J.

L., (2014). Consensus on circulatory shock and

hemodynamic monitoring. Task force of the European

Society of Intensive Care Medicine. Intensive care

medicine, 40(12), 1795–1815.

Asafuddoula, M., Shaheen, F. & Verma, B., (2016).

Impact of automatic feature extraction in deep

learning architecture. In 2016 International conference

on digital image computing: techniques and

applications (DICTA)(pp. 1-8). IEEE.

Badawi, O., Celi, L. A., Johnson, A. E., Mark, R. G.,

Pollard, T. J., & Raffa, J. D., (2018). The eicu

collaborative research database, a freely available

multi-center database for critical care research.

Scientific Data, 5(1).

https://doi.org/10.1038/sdata.2018.178

Bediako, Y., Djomkam Zune, A. L., Duodu, S, Nganyewo,

N. N., Olwal, C. O., Owoicho, O., &Tapela, K.,

(2021). Parallels in sepsis and covid-19 conditions:

Implications for managing severe covid-19. Frontiers

in Immunology, 12.

Chang, X., Chen, X., Huang, P.-Y., Li, Z., Ren, P., Wang,

X, & Xiao, Y., (2020). A survey of deep active

learning. arXiv.org.

Data: Badawi, O., Johnson, A., Johnson, A., & Raffa, J.,

(2021). eICU Collaborative Research Database Demo

(version 2.0.1). PhysioNet.

https://doi.org/10.13026/4mxk-na84.

Ej0Cl6/Deep-Active-Learning: Deep active learning.

GitHub. (n.d.). https://github.com/ej0cl6/deep-active-

learning.

Genga, K. R. & Russell, J. A., (2017). Update of sepsis in

the intensive care unit. Journal of Innate Immunity,

9(5), 441–455.

Sepsis and organ failure definitions and guidelines,

(1993). Critical Care Medicine, 21(10), 1612.

Shang, Y., & Wang, D., (2014). A new active labeling

method for deep learning. (2014) International Joint

Conference on Neural Networks (IJCNN).

Predictive Model of Septic Shock Staging Base on Continuing Invasive Hemodynamic Monitoring

523