Predictive Models of Ward Admissions from the Emergency

Department

Laiene Azkue

1,2 a

, Jon Kerexeta

1,3 b

, Jorge Sampedro

, Moisés Espejo

and Nekane Larburu

1,3 c

Vicomtech Foundation, Basque Research and Technology Alliance, (BRTA), 20009 Donostia, Spain

Biomedical Engineering Department, Mondragon Unibertsitatea, 20500 Mondragón, Spain

Biodonostia Health Research Institute, 20014 San Sebastián, Spain

Asunción Klinika, 20400 Tolosa, Spain

mespejo@clinicadelaasuncion.com

Keywords: Emergency Department, Ward Admission, Predictive Models, Machine Learning, Artificial Intelligence.

Abstract: The demand for emergency department (ED) care has increased significantly in recent years, mainly due to

factors such as the increase in chronic diseases, aging population and urban population growth. The large

influx of patients can lead to overcrowding and resource allocation problems, which impact the quality of

care. A new tool to improve patient severity classification systems could improve ED care and avoid

inappropriate admissions. Therefore, we propose the development of an artificial intelligence model to predict

ED ward admissions. The proposed model uses electronic medical records from the Asunción Klinika in Spain

and environmental data. Three models are created at different stages of ED: arrival model which predicts

admission upon patient arrival, triage model which predicts admission after clinicians’ triage and the last one,

laboratory model which make use of triage model data and laboratory analysis to estimate the risk among the

most critical patients. The arrival model achieved an AUC of 0.801, the triage model achieved an AUC of

0.854, and the laboratory model achieved an AUC of 0.781. These models provide valuable information for

efficient patient management and resource allocation in the ED, contributing to improved patient care and the

adequacy of hospital admissions.

1 INTRODUCTION

The demand for emergency department (ED) care has

increased considerably worldwide in recent years.

During the pandemic, it is evident that due to

COVID-19 there has been a disproportionate increase

in demand on EDs, leaving them overwhelmed

(WHS, 2023). However, regardless of the pandemic,

there is a considerable increase due to factors such as

an aging population, increase in chronic diseases, lack

of access to primary care, urban population growth,

and changes in lifestyles (McKenna et al.,

2019),(Lowthian et al., 2011) .

Hospital management is affected by the

increasing demand for ED care. The high influx of

patients can lead to overcrowding, long waiting times,

and challenges in resource allocation, which can

https://orcid.org/0009-0009-2266-7985

https://orcid.org/0000-0002-6516-8619

https://orcid.org/0000-0003-0248-7783

impact quality of care and patient satisfaction (Sun et

al., 2013). Rapid identification of the worst-off

patients to prioritise patient care could improve it.

Clinicians’ triage plays a crucial role in today's

EDs. It is a patient stratification system that allows

the identification and prioritization of patients

requiring immediate medical attention. The triage

evaluates parameters such as vital signs,

symptomatology, and initial clinical assessment

(Yancey and O’Rourke, 2023). This task involves the

intervention of the clinical caregivers, which can be

costly. Artificial intelligence (AI) can provide rapid

and improved evaluation when assessing patients

according to their severity.

To address this issue, we propose the

development of three AI models capable of predicting

ward admissions from the ED in three different

Azkue, L., Kerexeta, J., Sampedro, J., Espejo, M. and Larburu, N.

Predictive Models of Ward Admissions from the Emergency Department.

DOI: 10.5220/0012202700003657

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 277-284

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

277

sequential stages. Each stage requires different grades

of clinical involvement and clinical tests. Thus, the

models may detect high risk patients, which faster the

process without clinical involvement. These models

are based on various clinical factors, including the

patient's health status, diagnostic test results, severity

of the medical condition, and environmental factors

which differ in each stage.

Previously, we have conducted the following

systematic review, (Larburu et al., 2023), focusing on

predictive models for ED ward admission. This

review showed that logistic regression is the most

used algorithm, along with the gradient boosting

algorithm. The most frequently used variables are

triage and age. It is worth noting that all the reviewed

articles have an unbalanced nature. The systematic

review cover articles published between 2011 and

2022, with sample sizes ranging from 2,476 to

3,189,204 number of patients. The highest

performance was achieved by (Hong et al., 2018)

with the XGBoost algorithm, obtaining an AUC of

0.92 (95% CI 0.92-0.93).

The article is structured as follows: dataset section

summarizes the databases used and the general

characteristics of the database. Later, in the

methodology section, the models to be created and the

methodology carried out are exposed. Finally, the

results are presented, together with a discussion and

conclusions in the following two sections.

2 DATASET

Two types of data are used: electronic health records

(EHR), and environmental data.

EHR: The data used for analysis were obtained

from Asunción Klinika of Tolosa, Basque Country,

Spain. Once the study was approved by the

committee, the necessary data for the analysis were

provided. These retrospective data include vital signs,

laboratory results, details of performed tests,

demographic information, and more. The data covers

a period from January 1, 2004, to December 31, 2022,

encompassing a total of 284,503 emergency cases

involving 75,913 unique patients. It also includes the

target variable, which indicates whether the patient is

admitted to the ward or not. This variable is

unbalanced since 16.4% are admitted to the ward.

Environmental: Environmental data is extracted

from (Open-Meteo, 2023), an open-source weather

API service. This API provides access to

meteorological data from around the world. The data

come from various governmental meteorological

services and research organizations. The data covers

the period from January 1, 2004 to December 31,

2022 and they are from Tolosa. The environmental

data collected include the variables of temperature,

wind chill, solar radiation, precipitation, evaporation,

relative humidity, pressure, wind speed and gusts.

The two datasets are unified and a database with

196,659 instances and 255 predictor variables, since

EDs have been reduced to the year range 2010 to

2022, as triage was implemented in 2010. The target

variable is unbalanced, since the number of patients

not admitted to the ward is greater than the number of

patients admitted to the ward.

A total of 255 predictor variables can be classified

into the following categories (Table 1): demographic

(2), triage information (16), clinical and laboratory

findings (199), environmental (19) and others, that is,

uncategorized EHR variable (19).

Table 1: Description of predictor variables.

Quantity type

Demographic 2 -

Sex -

Categorical

Female 45.4%

Male 54.6%

Unknown 0.001%

Age (50.76±23.77) -

Numerical

-20 11.34%

20-40 24.05%

40-60 26.41%

+60 38.2%

Triage information 16 -

Triage -

Ordinal

categorical

0 – Non-urgent 0.006%

1 – Minor Urgent 0.95%

2 – Urgent 3.45%

3 – Emergency 18.51%

4 – Critical emergency 66.55%

5 – Immediate Life

Threatening

7.33%

NaN 3.2%

Vital signs Numerical

Clinical and laboratory

findings

199 -

Laboratory data 64 Numerical

tests performed 135 Categorical

Environmental 19 Numerical

Others 19 -

ED data 18

Numerical /

Categorical

Cause of attendance 1

Categorical

Common Disease 68.69%

Personal Accident 20.05%

Laboral Accident 5.7%

Sports Accident 2.13%

Traffic Accident 1.77%

School Accident 1.34%

Pregnancy 0.18%

Professional disease 0.134%

Undetermined 0.006%

HEALTHINF 2024 - 17th International Conference on Health Informatics

278

3 METHODS

This section discusses the methodology to be carried

out for the creation of predictive models. The

objective is the prediction of ward admissions from

the ED. For this purpose, with the data provided by

Asunción Klinika and acquired by Open-Meteo, three

models are created at three different stages of the ED

attendance (Figure 1):

Arrival Model: Model able to predict ward

admission from the ED upon patient arrival. This

model is trained with data prior to triage, such as

demographic data and environmental variable data.

Triage Model: Model capable of predicting ward

admission from the ED after triage. This model is

trained with data available at triage, such as,

demographic data, environmental variables data and

data obtained at triage: the triage value and vital signs

acquired at triage.

Laboratory Model: Model able to predict ward

admission from the ED for patients undergoing

laboratory tests after triage (smaller population). This

model is trained with post-laboratory test data, such

as demographic data, environmental variable data,

triage data and laboratory test data.

These three models are an efficient and fast way

to classify urgent patients from non-urgent patients

from the first moment of care. As new tests/analyses

are carried out in the ED, more accurate models will

be used as they make use of this new information.

First, on arrival at the ED, we will have a result of the

arrival model to be able to estimate the patient's risk.

Secondly, after clinicians triage the patient, we will

be able to obtain the updated risk with this new

variable (triage model). Finally, in case the patient

has not yet been discharged and laboratory tests have

been conducted, the risk will be updated with these

tests (laboratory model).

3.1 Data Preprocessing

To ensure the data is clean and suitable for predictive

modelling, several preprocessing steps are applied.

Initially, One-Hot Coding is performed to transform

the categorical variables into numerical ones.

Subsequently, missing values are imputed using the

KNN (K-Nearest Neighbors) method (Juna et al.,

2022). Finally, the classes of the target variable are

balanced, but only in the training database (70%)

using the random undersampling technique.

3.2 Feature Selection

The most related variables with the admission in ward

are estimated using a combination of the Boruta

method and the importance of the dependent feature

of the XGBoost model. On the one hand, the Boruta

method is a feature selection technique that helps to

identify the most relevant variables in a dataset

(Kursa and Rudnicki, 2010). On the other hand, the

importance of the dependent feature of the XGBoost

model allows to evaluate the degree of influence that

each variable has on the prediction of the model. The

variables selected by both techniques have been

selected as the relevant ones for the predictive models.

This methodology is carried out on the triage

model, since this is the model that has been given

more importance. This is because in the state of the

art it is the most studied model, that is, all models use

the triage variable and focus on the instant after

triage. Then, among these variables, the clinicians’

triage related variables are discarded for arrival model

training. Similarly, the laboratory model is trained

with the variables selected for the triage model, plus

variables acquired after triage.

3.3 Predictive Models

The predictive models implemented include the

following algorithms: Logistic Regression, K-NN,

Gaussian Naives Bayes, Bernoulli Naives Bayes,

Decision Tree Classification, Random Forest

Figure 1: Predictive system workflow for the emergency department.

Predictive Models of Ward Admissions from the Emergency Department

279

Classification, Gradient Boosting Classifier (GBC),

eXtreme Gradient Boosting (XGBoost), AdaBoost

(Adaptive), CatBoosting (Categorical), LightGBM

(Light Gradient Boosting Machine), MLP (Multilayer

Perceptron) (Juna, 2022; Theng, 2020; Kelleher,

2020; Natras, 2022).

3.4 Evaluation Methods

For model evaluation, the database is divided into

training (70%) and validation (30%). The best

hyperparameters are searched in the training set using

GridSearch. Finally, a 10-fold cross-validation is

used during the training phase (Browne, 2000).

The

evaluation metrics used to assess the model's

performance in both the validation set and during

cross-validation training are as follows (Hossin and

Sulaiman, 2015): confusion matrix, ROC-AUC,

accuracy, precision, recall and F1.

4 RESULTS

This section presents the results obtained from each

of the three models (see Figure 1), along with their

optimal hyperparameters.

4.1 Triage Model

The triage model is the model that predicts ward

admission after the clinical caregivers have

conducted the triage. Hence, it makes use of triage

information for the prediction.

4.1.1 Training Triage Model

After the feature importance (see Section 3.2), it was

observed that triage, cause of attendance, age, and sex

are the main variables.

Table 2: Result of algorithm XGBoost in the training of

model triage.

XGBoos

AUC 0.854(95% IC 0.849-0.858)

Accurac

0.773(95% IC 0.768-0.777)

Precision 0.757(95% IC 0.750-0.763)

Recall 0.803(95% IC 0.796-0.810)

F1 0.779(95% IC 0.775-0.784)

With these variables, predictive models were

created using the 10-fold cross validation technique

with 70% of the database.

XGBoost algorithm obtained the best results with an

AUC value of 0.854 (95% CI 0.849-0.858) (Table 2).

4.1.2 Triage Model Validation Results

Validation is performed with the XGBoost algorithm

at the default threshold of 0.5. The model is validated

with the remaining 30% of the database. This model

obtains an AUC of 0.858. This model classifies

74.29% of the negative class correctly, in addition to

classifying 81.22% of the positive class correctly.

Table 3 shows that the value of precision and F1

score decreases due to the imbalance of the validation

dataset.

Table 3: Triage model comparison in training and

validation.

Trainin

Validation

AUC 0.854 0.858

Accurac

0.773 0.755

Precision 0.757 0.384

Recall 0.803 0.812

F1 0.779 0.521

4.2 Arrival Model

The arrival model is the model that predicts ward

admissions at the time of arrival at the ED, before the

triage. Therefore, an estimate of the triage or patient

status could be made with the output of the arrival

model, since clinicians have not yet get involved and

this model can help them in their decision making.

4.2.1 Training Arrival Model

For this model we have used the same variables as in

the triage model, except from triage information,

which has been removed, since this model aims to

predict the risk of ward admission before the triage.

Hence, the cause of attendance, age and sex are solely

used. With these variables, predictive models were

created using the 10-fold cross validation technique

with 70% of the database.

The best performing model is Gradient boosting

with an AUC value of 0.801 (95% CI 0.796-0.806)

(Table 4).

Table 4: Result of algorithm Gradient Boosting in the

training of model arrival.

Gradient Boostin

AUC 0.801(95% IC 0.796-0.806)

Accurac

0.731(95% IC 0.726-0.735)

Precision 0.721(95% IC 0.716-0.725)

Recall 0.753(95% IC 0.748-0.757)

F1 0.736(95% IC 0.732-0.741)

HEALTHINF 2024 - 17th International Conference on Health Informatics

280

4.2.2 Arrival Model Validation Results

The Gradient Boosting model was validated with the

0.5 threshold. The model obtained an AUC value of

0.805. It predicts well 76.82% of patients admitted to

the ward and 70.45% of those not admitted to the

ward.

Table 5 shows the difference between the training

and validation results. In general, all the metrics have

maintained their training value in the validation,

except the accuracy, and therefore, the f1 score. This

is because the training dataset is balanced, and the

validation dataset is not.

Table 5: Arrival model comparison in training and

validation.

Trainin

Validation

AUC 0.801 0.805

Accurac

0.731 0.715

Precision 0.721 0.338

Recall 0.753 0.768

F1 0.736 0.469

4.3 Laboratory Model

Laboratory model is the model which predicts ward

admission from the ED for patients after triage and

with laboratory tests. The database used in this model

is reduced from 196,659 emergencies to 71,982

emergencies since those were the ones that had

laboratory information. This is the 36.6% from

previous cases. The imbalance of the data changes:

from the total number of these cases (n=71,982),

39.59% are admitted to the ward.

4.3.1 Training Laboratory Model

In this model we used triage variables, cause of

attendance, age, sex, laboratory data (red blood cells,

hemoglobin, hematocrit, MCV, MCH, MCHC, etc),

vital signs and environmental variables (temperature,

thermal sensation, precipitation, etc). The Gradient

boosting model obtains the best results with an AUC

value of 0.788 (95% CI 0.785-0.792) (Table 6).

Table 6: Result of algorithm Gradient Boosting in the

training of model laboratory.

Gradient Boostin

AUC 0.788 (95% IC 0.785-0.792)

Accurac

0.715 (95% IC 0.712-0.718)

Precision 0.719 (95% IC 0.716-0.723)

Recall 0.705 (95% IC 0.699-0.711)

F1 0.712 (95% IC 0.709-0.716)

4.3.2 Laboratory Model Validation Results

The selected model has been validated and the results

of the Table 7 are obtained. Considering that the

population has been reduced by discarding patients

who are initially in the best condition (n=71,982), it

is understandable that the predictive ability decreases

among the severe patients. Therefore, we have

applied the triage model in this reduced cohort to

compare whether the use of laboratory data improves

the predictive ability.

Figure 2: Confusion matrices (in %) to compare the triage

model (a) and the laboratory model (b) in the validation set

of the reduced data.

Table 7: Laboratory model comparison in training and

validation with the triage model in the reduced database.

Training Validation

Triage

model

AUC 0.781 0.779 0.728

Accurac

0.711 0.710 0.662

Precision 0.622 0.622 0.561

Recall 0.691 0.685 0.680

F1 0.654 0.652 0.615

The results of the laboratory model based on

laboratory data outperform both in validation and

training the results of the triage model in the reduced

database, as shown in the Table 7. A comparison of

the classification of laboratory model (b) and the

triage model in this cohort (a) can be seen in Figure 2

For positive predictions, laboratory model improves

by 1.08% over the triage model. In the case of

negative predictions, it improves by 7.33%, taking

into account that this class is the majority class.

5 DISCUSSION

Promising results were obtained using three

predictive models at different times during

emergency care: arrival model, triage model and

laboratory model. These three models predict ward

admission, but at different stages of Emergency

Predictive Models of Ward Admissions from the Emergency Department

281

Department (ED): arrival model predicts admission to

the ward on arrival of the ED, whereas the triage

model predicts it after triage has been performed.

Finally, laboratory model predicts ward admission for

patients undergoing laboratory tests, that is, it is a

model trained with a smaller population than the other

two models.

These models benefit healthcare personnel in

patient management by providing the ability to

evaluate ward admission at different points of care.

The accuracy of the models improves with an

increased number of tests conducted. Consequently,

the laboratory model demonstrates greater predictive

capacity compared to the arrival model. In addition,

the arrival model proves valuable in estimating

potential ward admissions prior to triage. Thus,

during periods of high demand, it may be possible to

create two care pathways: one with patients with a

high probability of admission and the other with

patients with a low probability of admission. This

approach could minimize the collapse in the ED and

enhance overall management. These models are

useful tools for the effective management of patients

according to their needs and can avoid unnecessary

admissions, improving not only the quality of care but

also patient safety. The following Table 8 shows the

results obtained for each model.

Table 8: Results of the ED workflow.

Nº Va

n AUC

Arrival model 3 196,659 0.805

Triage model 4 196,659 0.858

Laborator

model 55 71,982 0.779

There are 196,659 emergency room attendances.

The arrival model has obtained an AUC value of

0.805, with the use of 3 predictor variables (cause of

attendance, age and sex) and the Gradient Boosting

model.

Afterwards, the patient is triaged by nursing and

with this evaluation, another model (triage model) is

created to improve the capacity of the initial model

(arrival model). The triage model makes use of the

XGBoost algorithm and uses 4 variables, which are

the same as those used in the arrival model, together

with the triage value. An AUC value of 0.858 and a

precision of 0.440 is obtained. A low precision value

is obtained since the data are unbalanced (83.6%

negative class and 16.4% positive class). This implies

that the model has a tendency to predict the majority

(negative) class more frequently due to the higher

number of examples in the dataset. As a result, the

model accuracy is affected. But actually, the model is

able to correctly predict 72.70% of the positive class

and to incorrectly predict 18.13% of the negative

class. However, having the database unbalanced the

number of patients in the 18.13% of the negative class

is higher than the 72.70% of the positive class, which

lowers the accuracy result.

Finally, an additional model (laboratory model) is

developed with the aim of predicting the need for

ward admission for those patients undergoing

laboratory tests. It is important to note that these

patients, in general, present a more severe health

condition compared to those who do not undergo such

tests. Therefore, this model is trained with a reduced

database (n=71,982) with the gradient boosting

algorithm, of which 39.6% are admitted to the ward.

In this model the 4 variables of the previous arrival

model used together with laboratory results data, vital

signs data and environmental data, obtaining an AUC

value of 0.781. The AUC decreases with respect to

the triage model, but this model is not the best model

to compare whether the AUC has improved. This is

because the two models have been trained with

different databases, that is, the triage model has been

trained with a database of 196,659 urgencies,

including those in best condition, and the laboratory

model with a database of 71,982 urgencies, which are

supposed to be the most severe patients. Therefore,

these results have been compared with a new triage

model, which makes use of the same reduced

database, but only with triage model variables, i.e.

without laboratory tests information. As for the AUC

value, the value of this new triage model is 0.728 and

the laboratory model obtains a value of 0.781, thus

improving predictivity. This model does not greatly

increase the correct prediction of the positive class,

but it reduces the number of false positives.

6 CONCLUSION

Three machine learning models have been developed

to predict ward admissions from the ED at various

stages during the emergency care process. In these

three models triage, age and cause of care variables

are the most important ones in terms of prediction. In

addition, the best performing models are XGBoost

and Gradient boosting. The arrival model is useful for

identifying patients who may require ward admission

from the beginning of their ED care, without triage.

The triage model, with the inclusion of the triage

value, improves predictive ability later in the process.

On the other hand, laboratory model focuses on

patients undergoing laboratory tests, offering greater

predictive accuracy for this specific group.

HEALTHINF 2024 - 17th International Conference on Health Informatics

282

Regarding the results obtained and the results of

the systematic review by (Larburu et al., 2023) which

summarizes the articles dealing with predictive

models of ward admissions, it can be seen that even

without the clinicians triage variable, the arrival

model exceeds a quarter of the articles in the

systematic review, which make use of variables at

triage. Therefore, this initial model, achieves results

comparable with the literature but without the need of

clinician’s involvement.

In the case of the triage model, we observe that

three studies demonstrate superior results in terms of

AUC: 0.92 (Hong et al., 2018), 0.89 (Cusidó et al.,

2022), and 0.877 (Cameron et al., 2015). It is worth

noting that these three articles have been trained on

much larger databases, consisting of 560 thousand, 3

million, and 255 thousand instances, respectively. We

have identified six articles that achieve similar AUC

value (confidence intervals overlap), generally

characterized by a comparable database size (Sun et

al., 2011; Martinez et al., 2012; Zlotnik et al., 2016;

Graham et al., 2018; Lucke et al., 2018; De Hond et

al., 2021). Lastly, our model outperforms five

articles, all of which, except for one (more than a

million instances), make use of much smaller

databases (Noel et al., 2019; Parker et al., 2019; Brink

et al., 2020; Feretzakis, Karlis, et al., 2022;

Feretzakis, Sakagianni, et al., 2022).

Finally, it is important to note that the laboratory

model cannot be directly compared with the models

in the systematic review. This is because the

laboratory model is trained on a small population

consisting specifically of patients undergoing

laboratory tests in the ED, patient in worse condition.

These machine learning models offer an

opportunity to improve the management and

efficiency of emergency departments from a clinical

perspective. These models can help prioritize and

allocate resources more effectively, streamlining

floor admission processes and optimizing patient

care, as well as achieving more efficient management

of available resources, ensuring timely and

appropriate care for each patient, and thus improving

clinical outcomes in the ED setting. Additionally,

these models can play a relevant role in reducing

hospital admission inadequacy, which directly

translates into improvements in patient safety (Puig et

al., 2004).

However, it is important to mention some

limitations of the study. One important limitation to

consider is the applicability of the model to different

clinical contexts or settings. Since machine learning

models are trained with data specific to a particular

institution or context, it is possible that their

performance and predictive ability may be affected

when applied in other settings with different

characteristics. The variables used in the model may

be related to clinical practice and procedures specific

to the home institution, which could limit its

usefulness elsewhere where the relevant variables

may vary.

ACKNOWLEDGEMENTS

The authors would like to thank INVIZA-Asunción

Klinica and STT Systems for their support in the

INURGE project. This work has been funded by the

research project INURGE (ZL-2022/00571) of the

Basque Government’s HAZITEK programme from

the public agency SPRI.

ETHICAL COMMITEE

The study was conducted in accordance with the

Declaration of Helsinki and approved by the Research

Ethics Committee of the Gipuzkoa Health Area

(protocol code SAMURG-2022-01, 6 February

2023).

REFERENCES

Brink, A., Alsma, J., Brink, H. S. et al. (2020). Prediction

admission in the older population in the Emergency

Department: the CLEARED tool. Neth J Med 78, 357–

367

Browne, M. W. (2000). Cross-Validation Methods. J Math

Psychol 44, 108–132. https://doi.org/10.1006/jmps.19

99.1279.

Cameron, A., Rodgers, K., Ireland, A. et al. (2015). A

simple tool to predict admission at the time of triage.

Emerg Med J 32, 174–179

Cusidó, J., Comalrena, J., Alavi, H. et al. (2022). Predicting

Hospital Admissions to Reduce Crowding in the

Emergency Departments. Appl Sci 12, 10764

De Hond, A., Raven, W., et al. (2021). Machine learning

for developing a prediction model of hospital admission

of emergency department patients: Hype or hope? Int J

Med Inf 152, 104496

Feretzakis, G., Karlis, G., et al. (2022). Using machine

learning techniques to predict hospital admission at the

emergency department. J Crit Care Med 8, 107–116

Feretzakis, G., Sakagianni, A., Loupelis, E. et al. (2022).

Predicting Hospital Admission for Emergency

Department Patients: A Machine Learning Approach.

Stud Health Technol Inf, 297–300

Predictive Models of Ward Admissions from the Emergency Department

283

Graham, B., Bond, R., et al. (2018). Using data mining to

predict hospital admissions from the emergency

department. IEEE Access 6, 10458–10469

Hong, W. S., Haimovich, A. D. and Taylor, R. A. (2018).

Predicting hospital admission at emergency department

triage using machine learning. PloS One 13, e0201016

Hossin, M. and Sulaiman, M. N. (2015). A review on

evaluation metrics for data classification evaluations.

Int J Data Min Knowl Manag Process 5, 1

Juna, A., Umer, M., Sadiq, S. et al. (2022). Water quality

prediction using KNN imputer and multilayer

perceptron. Water 14, 2592

Kelleher, J. D., Namee, B. M. and D’Arcy, A. (2020).

Fundamentals of Machine Learning for Predictive

Data Analytics, second edition: Algorithms, Worked

Examples, and Case Studies. MIT Press.

Kursa, M. B. and Rudnicki, W. R. (2010). Feature Selection

with the Boruta Package. J Stat Softw 36, 1–13.

https://doi.org/10.18637/jss.v036.i11.

Larburu, N., Azkue, L. and Kerexeta, J. (2023). Predicting

Hospital Ward Admission from the Emergency

Department: A Systematic Review. J Pers Med 13, 849.

https://doi.org/10.3390/jpm13050849.

Lowthian, J. A., Curtis, A. J., Cameron, P. A. et al. (2011).

Systematic review of trends in emergency department

attendances: an Australian perspective. Emerg Med J

28, 373–377. https://doi.org/10.1136/emj.2010.099226.

Lucke, J. A., de Gelder, J., et al. (2018). Early prediction of

hospital admission for emergency department patients:

a comparison between patients younger or older than 70

years. Emerg Med J 35, 18–27

Martinez, E., Fernández, C., et al. (2012). Prediction model

for in-hospital admission in patients arriving in the

emergency department. In Anales del Sistema Sanitario

de Navarra (207–217)

McKenna, P., Heslin, S. M., et al. (2019). Emergency

department and hospital crowding: causes,

consequences, and cures. Clin Exp Emerg Med 6, 189–

195. https://doi.org/10.15441/ceem.18.022.

Natras, R., Soja, B. et.al. (2022). Ensemble Machine

Learning of Random Forest, AdaBoost and XGBoost

for Vertical Total Electron Content Forecasting.

Remote Sens 14, 3547. https://doi.org/10.3390/

rs14153547.

Noel, G., Bonte, N., et al. (2019). Real-time estimation of

inpatient beds required in emergency departments. Eur

J Emerg Med 26, 440–445

Open-Meteo (2023). Open-Meteo. Available at:

https://openmeteo.substack.com/embed [Accessed

December 22, 2023].

Parker, C. A., Liu, N., Wu, S. X. et al. (2019). Predicting

hospital admission at the emergency department triage:

A novel prediction model. Am J Emerg Med 37, 1498–

1504

Puig, C., Carrasco, G., et al. (2004). Factores asociados a la

inadecuación de los ingresos por urgencias en un

hospital privado de tercer nivel. Rev Calid Asist 19,

363–369. https://doi.org/10.1016/S1134-282X(04)777

25-8.

Sun, B. C., Hsia, R. Y., et al. (2013). Effect of emergency

department crowding on outcomes of admitted patients.

Ann Emerg Med 61, 605-611.e6. https://doi.org/10.1016

Sun, Y., Heng, B. H., Tay, S. Y. et al. (2011). Predicting

hospital admissions at emergency department triage

using routine administrative data. Acad Emerg Med 18,

844–850

Theng, M. and Theng, D. (2020). Machine Learning

Algorithms for Predictive Analytics: A Review and New

Perspectives. https://doi.org/10.37896/HTL26.06/1159.

WHS (2023). World health statistics 2023: monitoring

health for the SDGs, sustainable development goals.

Yancey, C. C. and O’Rourke, M. C. (2023). Emergency

Department Triage. In StatPearls. Treasure Island (FL):

StatPearls Publishing. Available at: http://www.

ncbi.nlm.nih.gov/books/NBK557583/ [Accessed June

23, 2023].

Zlotnik, A., Alfaro, M. C., et al. (2016). Building a decision

support system for inpatient admission prediction with

the Manchester triage system and administrative check-

in variables. CIN Comput Inform Nurs 34, 224–230

HEALTHINF 2024 - 17th International Conference on Health Informatics

284