International Journal of Artificial Intelligence and Applications (IJAIA), Vol.15, No.

3, May 2024

CLASSIFYING EMERGENCY PATIENTS INTO

FAST-TRACK AND COMPLEX CASES USING
MACHINE LEARNING
Ala’ Karajeh1 and Rasit Eskicioglu2
1
Biomedical Engineering Program, University of Manitoba, Winnipeg, Canada
2
Department of Computer Science, University of Manitoba, Winnipeg, Canada

ABSTRACT
Emergency medicine is a lifeline specialty at hospitals that patients head to for various reasons, including
serious health problems, traumas, and adventitious conditions. Emergency departments are restricted to
limited resources and personnel, complicating the optimal handling of all received cases. Therefore,
crowded waiting areas and long waiting durations result. In this research, the databases of MIMIC-IV-ED
and MIMIC-IV were utilized to obtain records of patients who visited the Beth Israel Deaconess Medical
Center in the USA. Triage data, dispositions, and length of stay of these individuals were extracted.
Subsequently, the urgency of these cases was inferred based on standards stated in the literature and
followed in developed countries. A comparative framework using four different machine learning
algorithms besides a reference model was developed to classify these patients into complex and fast–track
categories. Moreover, the relative importance of employed predictors was determined. This study proposes
an approach to deal with non-urgent visits and lower overall waiting times at the emergency by utilizing
the powers of machine learning to identify high-severity and low-severity patients. Given the provision of
the required resources, the proposed classification would help improve the overall throughput and patient
satisfaction.

KEYWORDS
Emergency Medicine, Triage Enhancement by Machine Learning, Emergency Patients Classification,
Identifying Fast-Track Patients. Identifying Severity of Emergency Cases.

1. INTRODUCTION
Patients visit the emergency department of a hospital for various reasons. While some individuals
visit emergency departments based on a personal perception of the need to be examined by a
physician, the scarcity of a nearby primary care facility, or following a friend's suggestion, others
may be referred to the emergency by a healthcare provider [1]. As a result, nurses and physicians
in emergency medicine usually treat various conditions ranging from life-threatening, such as
strokes and heart attacks, to minor ones, including localized pain and obtaining a medication
refill. Patients who head to the emergency expect to attain efficacious patronage, provided
without delays and reachable in alignment with their locations and schedules [2].

Several systems were established in developed countries to guide healthcare professionals on

triaging received cases based on the level of their acuity, such as the Emergency Severity Index
(ESI) and the Australasian Triage Scale [3][4]. However, previous studies pointed to constraints
in emergency triage, including the inconsistency of staff performance and limited accuracy of
assessment, which indicate that existing triage approaches efficiency is insufficient, including
their abilities to identify lower-severity cases that can subsequently be fast-tracked [5][6].

DOI:10.5121/ijaia.2024.15305 53
 International Journal of Artificial Intelligence and Applications (IJAIA), Vol.15, No.3, May 2024
Moreover, a few obstacles manifested in emergency units, particularly long waiting times and
being overloaded by cases exceeding the available personnel and resources [7][8]. Various
factors lead to such impediments, including unsound triage appraisal besides non-urgent
emergency visits, which exerts extra pressure on personnel at emergency departments [9]. A
significant proportion of emergency visits in the USA were comprised of such non-critical visits
[10].

Consequently, these issues negatively reflect on the level of care provided to received patients
and impact the satisfaction level of these people as they must wait for a longer time to be
examined and treated at the emergency room [7][8]. Artificial intelligence (AI) technologies are
changing the scenes in several fields. They seem promising in healthcare, where they can be
implemented in different use cases, including data extraction, diagnostics, prognostics, and
development of drugs [11][12].

In emergency medicine, most previous studies focus on predicting outcomes and length of stay
rather than improving the triage process itself [13]. While Chang et al. utilized tree-based
algorithms and boosting methods to identify low-severity cases among level III patients in
Taiwan, they excluded levels IV and V in their study. Moreover, they did not incorporate neural
network algorithms, which are prominent in different classification scenarios [14]. To our
knowledge, no study has explored categorizing emergency patients comprehensively into fast-
track and complex cases based on the composite of the length of stay, disposition, and triage
score outcomes by implementing machine learning (ML) algorithms, including artificial neural
networks.

Building on these considerations, this research proposes a ML-based model that can classify
emergency cases thoroughly into fast-track and complex cases. This research proposes four
different classifying models, including a neural network model, and compares these models to a
reference one. The models were developed from a clinical dataset of a medical center that
implements the ESI system for emergency triage scores. Cases assigned to levels III, IV, and V
were eventually discharged after spending less than four hours in the emergency division, and
they were considered fast-track cases. As a result, they require fewer resources from emergency
care providers due to being less critical. On the contrary, complex cases necessitate a different
handling approach since they require more attention and resources. This paper starts with an
abstract and an introduction, followed by the research methodology. Then, the results are
elaborated and discussed thoroughly. Finally, the paper is wrapped up with conclusions and
relevant future work.

2. METHODS
2.1. Aim, Design and Setting of the Study

This study aimed to harness healthcare big data along AI’s exceptional pattern recognition and
predictive capabilities to lay the foundation for optimizing workflow in emergency departments.
Therefore, it aimed to develop a model based on ML approaches that can help identify low-
severity cases besides demanding ones among emergency patients. Furthermore, it purported to
verify the performance of the developed model on a subset that was not engaged in the training
phase as well as to present an illustration of how the predictors function regarding the
classification outcome.

To accomplish the objectives of this research, a clinical dataset containing patient information
that is collected at the time of triage in the emergency was sought. Moreover, past health records

54
 International Journal of Artificial Intelligence and Applications (IJAIA), Vol.15, No.3, May 2024
concerning demographic details and registration information for the same individuals were
required to obtain a broader perspective of the studied population. Preferably, retrospective data
from a healthcare institution that implements a renowned triage system, such as the Canadian
Triage and the Acuity Scale and Emergency Severity Index scale, were desired where they are
utilized widely worldwide. Such systems guide emergency nurses and providers on the acuity
level of received cases, considering their physiological status, perceived pain and reason for the
visit.

Figure 1: Methodology steps followed in this research.

To carry out this research, two clinical databases were harnessed, which had retrospective data
pertaining to patients who visited the emergency department of Beth Israel Deaconess Medical
Center in the USA for the period from 2011 to 2019. The respective databases were MIMIC-IV-
ED (v2.0) and MIMIC-IV (v2.0), where the first one included the information gathered at
emergency, including triage score and vital signs. On the other hand, demographic data were
debriefed from the latter one, such as age and race. The respective health institution is a tertiary
academic medical center which provides healthcare and research services besides being involved
in Academia and teaching in affiliation with Harvard Medical School. The center is in Boston,
MA where it has approximately 750 beds and receives approximately 50,000 emergency visits
yearly.

2.2. Characteristics of the Participants

The records extracted from the two databases were for patients aged 18 and older. The conditions
varied largely from simple grazes and medicine refills to life-threatening heart disorders. Clinical
symptoms of the reported visits exceeded 60,000, which were entered as free texts, while some of
them included two or three different symptoms. Cases with unavailable triage scores were
omitted from the dataset, while cases with missing vital signs or pain values were replaced so that
we could have a greater deal of samples available for training and testing of the ML models. A

55
 International Journal of Artificial Intelligence and Applications (IJAIA), Vol.15, No.3, May 2024
considerable portion of the documented cases were assigned to the acuity levels of two and three
by the triage nurse at the time of assessment.

2.3. Dataset Access and Pre-Processing

Accessing the data was carried out through the website of Physionet.org. Subsequently, they
requested completing specific requirements to grant access to both databases: signing up as a
credentialed user, signing a consent regarding data usage, and finishing an online human ethics-
focused course. The requested data became available accordingly.

Data analysis and machine-learning models were developed using R language software (v 4.2.1)
besides the RStudio (v2023.06.1 + 524) for Windows 11. Three different datasets files were
employed: patients.csv, triage.csv, and edstays.csv. The first dataset was drawn from the MIMIC-
IV (v2.0), while the other two datasets were pulled from the MIMIC-IV-ED (v2.0). Hence, the
three files were combined where first patients.csv and triage.csv were incorporated by subject_id.
Then, the edstays file was merged with the consequent file according to the identifiers: stay_id
and subject_id.

2.4. Dataset Processing

Figure 2: Data pre-processing and processing steps.

Following the merging step, specific columns were omitted. As a result, this research excluded
the columns of date of birth (dod) along with hospital admission identifier (hadm_id) for non-
relevancy. The same process was applied to the duplicate gender column. Next, in-time and out-
time data points were employed to extract the year of the emergency visit. After that, the age of
the recorded cases was concluded with the assistance of the in_Year variable according to the
following equation: age = in_Year−anchor_year+anchor_age. Moreover, the length of stay for
documented cases was calculated by subtracting the intime from the outtime column. It was also
converted into hours format and rounded to two decimal places. The names of columns were
amended to facilitate the handling of variables during the coding process.

A few modifications were made to the dataset where categorical variables were transmuted into
factors, and the Pain and the Acuity were made as numerical variables to facilitate data handling
and feature extraction. Then, NA’s values were replaced using the K-Nearest Neighbors method,
which imputes such values using the closest observations in the dataset, replacing these values
more accurately. The identifiers of both stay and patient were removed. In addition, the race and
in-year variables were excluded from the dataset to avoid any potential bias in the model due to
the former and the latter did not reflect the actual year of emergency visit. Further, a classifying
column was created to group patients into Fast-track and Complex. Cases with a length of stay
shorter than four hours in the emergency were considered fast-track, which is considered the
standard length of stay for emergency patients and applied in different developed countries such

56
 International Journal of Artificial Intelligence and Applications (IJAIA), Vol.15, No.3, May 2024
as Canada and Australia [15][16]. Furthermore, incorporated cases should be assigned to acuity
levels III, IV, or V and discharged at the end of the emergency visit. The categorization was
carried out using the Dplyr library’s functions. The Disposition, Acuity and Length of Stay
columns were omitted after that as they were partly involved in inferring the two categories.

Major accompanying symptoms for the fast-track cases were identified to provide some insights
into the characteristics of these patients. Furthermore, to narrow down the tremendous variety of
complaints in the dataset. As a result, the most frequent ten complaints and the top thirty
corresponding from the subgroup of fast-track – level III and the subgroup of fast-track – level IV
& V were utilized to bring about a more balanced dataset. The Forcats library’s functions were
applied to combine similar symptoms. Next, the outliers of numerical predictors were converted
into NA’s and replaced using K-Nearest Neighbors method. Following the outliers handling, the
remaining data were centered and scaled to lay the ground for effective classifying models.

2.5. Model Development

A reference model was established based on the variables of vital signs, pain, chief complaints,
arrival mean and age to act as a baseline model in this research. Logistic regression (LR) was
selected for this purpose, considering the simplicity of this algorithm and the nature of the
outcome as a binary one. The Caret package was utilized for its variety of algorithms to establish
ML-based classifying models. Four different models were created by using the algorithms of
Random Forests (RF), Stochastic Gradient Boosting (SGB), Multi-Layer Perceptrons (MLP), and
Classification and Regression Trees (CART) [17][18][19]. Prior to the training step, a data split
was carried out where 80% of the data set was assigned to the training set, and the rest was
allocated for the testing procedure. Furthermore, the predictor column was excluded from the
training set. Moreover, the five-fold cross-validation technique, which encompasses splitting the
data into five similar parts, where four are utilized for training while the remaining part is held
for validation, was incorporated into the training process to produce more generalizable
predictive models. This process is repeated for the five parts, and the model performance is
computed as an average of the five predictions. One hot encoding technique was also applied to
the categorical variables of the dataset to convert them into numerical ones. In turn, it facilitates
running the algorithms of MLP and CART. For reproducibility purposes, two seeds were set to
values 12345 and 1234 for data segmentation and algorithm training, respectively.

2.6. Model Tuning

Hyperparameters of the four employed algorithms were adjusted prior to the validation stage.
Hyperparameters refer to the parameters through which we can control how an algorithm runs
and learns an assigned task. Consequently, various parameters were passed into the Tunegrid to
optimize the resulting models' performance. These parameters included the number of hidden
layers and neurons for the MLP, complexity parameter for the CART algorithm, interaction depth
and number of trees for the SGB algorithm, besides minimum node size and number of features
sampled at a split for the RF algorithm. The activation function that was employed in the MLP
algorithm is the logistic regression function. Moreover, the number of hidden layers that were
experimented with was from one to three layers during the tuning phase. Node numbers were
increased gradually, starting with one node in a single layer and concluding with fifteen nodes in
the whole for the three hidden layers, taking into consideration the computational cost of having
multiple hidden layers for such models. The below figure illustrates the basic structure of an
MLP network.

57
 International Journal of Artificial Intelligence and Applications (IJAIA), Vol.15, No.3, May 2024

Figure 3: Schematic representation of an MLP network.

2.7. Performance Evaluation of the Model

After completing the development phase, the model was set to classify the new data in the test
set. Then, these outcomes were compared to the actual ones, and performance measures of area
under the curve (AUC), sensitivity and precision were computed to assess how well the models’
identified cases. AUC, which is the area under the receiver operating curve, represents the overall
performance of the predictive model, while sensitivity and precision evaluate the model’s
capability for fast-track case detection, where sensitivity refers to the probability of predicting a
positive outcome when the actual one is positive, and precision is the ratio of correctly predicted
positive instances to the total predicted positive instances. Confidence intervals were determined
by applying the DeLong method to compare statistically the AUC of the models with respect to
the reference model. Evaluation metrics were calculated for each of the five models. To provide
some insight into features that are significant for the developed models, the Variables' importance
was determined by employing relative predictors' importance attribute of the SGB model.

3. RESULTS
After combining both required datasets, the resulting dataset had 400,443 complete cases. The
merged dataset had the needed specifics, including vital signs readings at triage, demographic
characteristics of the received patients, length of stay of the visits, and the ultimate endpoints at
emergency for the respective individuals. Further information on the attributes of the utilized
dataset is demonstrated in the table below.

58
 International Journal of Artificial Intelligence and Applications (IJAIA), Vol.15, No.3, May 2024
Table 1: Demographic and triage characteristics of the merged dataset.

Variable Mean ± SD/Count

Temperature (°F) 98.01±3.75
Heart rate (Beat per minute) 84.91±17.62
Respiratory rate (Breath per minute) 17.55±23.02
Oxygen saturation (%) 98.51±16.86
Systolic Blood Pressure (mmHg) 135.17±47.60
Diastolic blood pressure (mmHg) 81.10±1058.99
Pain (0-10 scale) 4.42±4.05
Age (year) 52.31±20.46
Acuity level
Level 1 13107
Level 2 132681
Level 3 224739
Level 4 28827
Level 5 1089
Chief complaint
Abdominal pain 17810 (4.45%)
Chest pain 13479 (3.37%)
Dyspnea 6711 (1.68%)
Status post fall 6676 (1.67%)
Headache 5143 (1.28%)
Gender
Male 181279
Female 219164
Race
White 215392
Black/African American 73888
Other 19568
Hispanic/Latino - Puerto Rican 13760
White - Other European 8724
Arrival Mean
Ambulance 137892
Helicopter 68
Other 1225
Unknown 11259
Walk in 249999

The subset derived from the primary dataset had 104,014 records and 12 variables, including the
outcome. These rows comprised 79,048 complex cases and 24,966 fast-track cases. Additionally,
the chief complaints of these patients constituted 42 manifestations. After replacing the outliers
using the K-Nearest Neighbor imputation, the observations’ count remained the same. The
following figure illustrates the two outcome groups and their corresponding number of cases,
stratified by the acuity scores.

59
 International Journal of Artificial Intelligence and Applications (IJAIA), Vol.15, No.3, May 2024

Figure 4: Distribution of outcome classes in the subset stratified by acuity levels.

Following the construction of the classifying models, the importance of predictors was attained
and scaled to be from zero to one hundred. The complaint came first as the most crucial factor in
the prediction process, followed by arrival transport mean and age. The least influential factors
in terms of contribution to the predicted response were diastolic blood pressure, respiratory rate,
and gender, respectively. The descending order of the influential variables is indicated in the
below figure.

Figure 5: Ranking of relative importance of predictors according to the SGB model.

The performance of the predictive models varied when they were validated during the testing
phase; the MLP algorithm statistically outperformed the reference model by scoring an AUC of
0.7594 (95%CI: 0.7523-0.7665), where it correctly classified the highest percentage of fast-track
cases among the five models. However, the other models could not surpass statistically the
reference model in terms of AUC. The following table demonstrates the performance assessment
for the generated ML models and the reference model followed by a figure for the receiver
operating curves of the five models:

60
 International Journal of Artificial Intelligence and Applications (IJAIA), Vol.15, No.3, May 2024
Table 2: Comparison of the Performance of the classifying models.

Model AUC (95% CI) Sensitivity (95% CI) Precision (95% CI)
MLP 0.7594 (0.7523-0.7665) 0.6399 (0.6264-0.6532) 0.6252 (0.6118-0.6385)
CART 0.7177 (0.7104-0.7249) 0.5185 (0.5046-0.5325) 0.6632 (0.6481-0.6780)
RF 0.7223 (0.7151-0.7296) 0.5245 (0.5106-0.5385) 0.6747 (0.6597-0.6894)
SGB 0.7309 (0.7237-0.7381) 0.5394 (0.5254-0.5533) 0.6872 (0.6724-0.7017)
LR (Reference 0.7221 (0.7149-0.7294) 0.5199 (0.5060-0.5339) 0.6846 (0.6695-0.6994)
Model)

Figure 6: Receiver operating curves of the developed ML models and the reference model.

4. DISCUSSION
This research aimed to devise an enhancing strategy so that the flow of care inside emergency
departments is expedited for stable cases – urgent and non-urgent by harnessing the ML
algorithms' predictive capacities. This method, in turn, would reflect on the contentment of
patients as it may make for receiving efficacious treatment within shorter periods. As a result, this
study presents four ML-based classification models besides a reference model developed from
two retrospective patient databases of the same healthcare facility, besides providing some
insights on the most significant factors for such classification. In this study, the MLP algorithm,
which had two hidden layers of nine and seven nodes, respectively, yielded leading outcomes
compared to the reference model.

MLP algorithm is a multi-layer neural network that comprises an input layer, an output layer, and
a variable number of hidden layers. Initial weights are set to the inputs and are utilized to predict
the outcome. This prediction is followed by iterative weight adjustments to decrease the mean
error and optimize the accuracy of the predicted results. The superiority of the MLP model
compared to the reference model could be attributed to the presence of hidden layers in its
network structure, which enables the neural network to learn complex tasks, including composite
health outcomes predictions. The classification outcome of this study was inferred based on the
acuity score, disposition and length of stay, which in turn might have limited the performance of
the other algorithms utilized in this study due to the complexity of this outcome. Concerning the
most influential predictors, the leading position gained by complaints could indicate how
significant the ailment patients have, whether in triaging them or appending them into the fast-
track cases lane. Individuals with less severe symptoms are more likely to be identified as fast-
61
 International Journal of Artificial Intelligence and Applications (IJAIA), Vol.15, No.3, May 2024
track patients. The second factor on the list was arrival mean, where a substantial portion of the
Complex group arrived by ambulance to the emergency department. Age predictor, which came
third, also plays a significant role in identifying the complexity of the received patient. The subset
revealed a significant difference between the suggested complex and fast-track categories, with
age specifics of 52.98±20.41 and 41±17.82, respectively.

Integrating such models into clinical decision-aid tools would help in the data-driven
classification of received cases based on their likely outcomes in terms of length of stay at the
emergency and the conceivable destination at the end of their visit. Implementing these tools and
setting up the required resources for faster emergency care for low-severity cases would reflect
on patients' waiting time and satisfaction with the service catered to them. Fast-tracking stable
patients in an emergency is a concept that has been investigated in a few studies, which in turn
indicated favorable consequences. Chrusciel et al. found that recognizing minor emergency cases
gave rise to a drop in the overall length of stay for patients. Moreover, it reduced the rate of cases
who departed the emergency before being examined by a provider [8]. The impact of this strategy
on older patients’ health was researched by Gasperini et al., where the fast-track group
demonstrated a significant drop in length of stay compared to the control group, besides a faster
discharge following the end of physician’s inspection [20]. However, harnessing the emerging AI
technologies for this purpose is hardly examined in the existing literature since there was only
one study on this topic for a dataset in Taiwan. Chang et al. suggested differentiating the low-
severity patients assigned to level III at the emergency by utilizing five ML algorithms, including
the XGBoost and the CatBoost. They established their model based on data from two health
institutions that employ the TTAS triage system in Taiwan [14]. Another study in Iran inspected
the possibility of navigating the workflow of emergency cases through simulation modelling and
ML algorithms, where they evaluated the impact of a few factors relevant to resources, numbers
of providers, and available inpatient beds on the waiting time of triage-run units and fast-track
units. Consequently, their assessment found a prominent enhancement for both outcomes [21].

Despite the presence of urgent care units in some developed countries, for instance, Canada, their
impact on relieving some of the pressure put on emergency may be limited due to the absence of
a comprehensive understanding of the differences between the two divisions and where to head in
case of unexpected illnesses [22]. This phenomenon aggravates the situation in societies
challenged by immigration waves and homelessness as these individuals tend to struggle to select
the proper healthcare facility that they should visit in case of ailments, preferring emergency care
over other healthcare services, which in turn exerts additional pressure on emergency
departments [23]. A previous study in the UK indicated that non-urgent visits to emergency
departments are more common among adults and young adults compared to seniors [24]. Hence,
these factors may hinder providing care to high-severity cases in emergency units as well as
vulnerable patients, for instance, the elderly population. In light of such challenges, streamlining
cases at emergency units using cutting-edge technologies, including artificial intelligence
systems, would help improve the flow of patients in emergencies. In turn, it could assist in better
serving both severe and non-severe cases, provided that the needed resources are assigned.

5. CONCLUSIONS AND FUTURE WORK

This study proposes classifying emergency patients into complex and fast-track. The indicated
categorization was based on utilizing ML algorithms and harnessing retrospective emergency
data. As a result, the results of testing four ML-based models besides a reference model on a
validation set that was not engaged in the development phase were presented. MLP model
statistically demonstrated a better performance compared to the reference model. Moreover, the
significance of the predictors and how they contribute to the outcome of predictions was
determined from the SGB model. Considering the inadequate performance of current triage
62
 International Journal of Artificial Intelligence and Applications (IJAIA), Vol.15, No.3, May 2024
approaches, Implementing ML algorithms in emergency departments may help identify high-
severity and low-severity cases in an emergency based on arrival means, vital signs, complaints,
and age besides pain. Consequently, it would help mitigate left-without-being-seen rates and
diminish waiting times for low-severity cases, given it is integrated into a decision aid tool and
applied along the required resources for rapid care of such categories, positively reflecting on
patient satisfaction with provided services and overall emergency throughput.

Since the Caret package provides limited hyperparameter tuning for neural networks through the
methods utilized, other packages that offer more extensive tuning will be explored in the future.
Moreover, deep learning models will be established and compared to the results of this paper.
The source code employed in this research can be accessed at [25].

REFERENCES

[1] A. O’Cathain, J. Connell, J. Long, and J. Coster, “‘clinically unnecessary’ use of emergency and
urgent care: A realist review of patients’ decision making,” Health Expectations, vol. 23, no. 1, pp.
19–40, Oct. 2019. doi:10.1111/hex.12995
[2] K. Hansen et al., “Updated framework on quality and safety in emergency medicine,” Emergency
Medicine Journal, vol. 37, no. 7, pp. 437–442, May 2020. doi:10.1136/emermed-2019-209290
[3] S. Bhaumik et al., “Prehospital triage tools across the world: A scoping review of the published
literature,” Scandinavian Journal of Trauma, Resuscitation and Emergency Medicine, vol. 30, no. 1,
Apr. 2022. doi:10.1186/s13049-022-01019-z
[4] J. M. Zachariasse et al., “Performance of Triage Systems in emergency care: A systematic review
and meta-analysis,” BMJ Open, vol. 9, no. 5, May 2019. doi:10.1136/bmjopen-2018-026471
[5] H. L. Tam, S. F. Chung, and C. K. Lou, “A review of triage accuracy and future direction,” BMC
Emergency Medicine, vol. 18, no. 1, Dec. 2018. doi:10.1186/s12873-018-0215-0
[6] J. S. Hinson et al., “Triage performance in emergency medicine: A systematic review,” Annals of
Emergency Medicine, vol. 74, no. 1, pp. 140–152, Jul. 2019.
doi:10.1016/j.annemergmed.2018.09.022
[7] K. B. Ahsan, M. R. Alam, D. G. Morel, and M. A. Karim, “Emergency department resource
optimisation for improved performance: A Review,” Journal of Industrial Engineering International,
vol. 15, no. S1, pp. 253–266, Nov. 2019. doi:10.1007/s40092-019-00335-x
[8] J. Chrusciel et al., “Impact of the implementation of a fast-track on emergency department length of
stay and quality of care indicators in the Champagne-ardenne region: A before–after study,” BMJ
Open, vol. 9, no. 6, Jun. 2019. doi:10.1136/bmjopen-2018-026200
[9] C. O’Keeffe, S. Mason, R. Jacques, and J. Nicholl, “Characterising non-urgent users of the
Emergency Department (ED): A retrospective analysis of routine Ed Data,” PLOS ONE, vol. 13,
no. 2, Feb. 2018. doi:10.1371/journal.pone.0192855
[10] S. S. Al-Otmy, A. Z. Abduljabbar, R. M. Al-Raddadi, and F. Farahat, “Factors associated with non-
urgent visits to the emergency department in a tertiary care centre, western Saudi Arabia: Cross-
sectional study,” BMJ Open, vol. 10, no. 10, Oct. 2020. doi:10.1136/bmjopen-2019-035951
[11] S. Islam. "Artificial Intelligence in Healthcare," International Journal of Engineering Materials and
Manufacture, vol. 6, no. 4, pp. 319–323, 2021.
[12] S. Secinaro, D. Calandra, A. Secinaro, V. Muthurangu, and P. Biancone, “The role of Artificial
Intelligence in healthcare: A structured literature review,” BMC Medical Informatics and Decision
Making, vol. 21, no. 1, Apr. 2021. doi:10.1186/s12911-021-01488-9
[13] G. Chenais, E. Lagarde, and C. Gil-Jardiné, “Artificial Intelligence in emergency medicine:
Viewpoint of current applications and foreseeable opportunities and challenges,” Journal of Medical
Internet Research, vol. 25, May 2023. doi:10.2196/40031
[14] Y.-H. Chang et al., “Machine learning–based triage to identify low-severity patients with a short
discharge length of stay in emergency department,” BMC Emergency Medicine, vol. 22, no. 1, May
2022. doi:10.1186/s12873-022-00632-6
[15] T. Tenbensel et al., “New Zealand’s emergency department target – did it reduce ed length of stay,
and if so, how and when?,” BMC Health Services Research, vol. 17, no. 1, Sep. 2017.
doi:10.1186/s12913-017-2617-1

63
 International Journal of Artificial Intelligence and Applications (IJAIA), Vol.15, No.3, May 2024
[16] I. Ashkenazi, L. Gefen, O. Hochman, and E. Tannous, “The 4-hour target in the emergency
department, in-hospital mortality, and length of hospitalization: A single center-retrospective
study,” The American Journal of Emergency Medicine, vol. 47, pp. 95–100, Sep. 2021.
doi:10.1016/j.ajem.2021.03.049
[17] A. Singh, N. Thakur, and A. Sharma, ‘A Review of Supervised Machine Learning
Algorithms’, 2016 International Conference on Computing for Sustainable Global Development
(INDIACom), pp. 1310–1315, 2016.
[18] R. Firmansyah, E. Utami, and E. Pramono, “Evaluation of naive bayes, Random Forest and
stochastic gradient boosting algorithm on ddos attack detection,” International Conference on
Information Science and Technology Innovation (ICoSTEC), vol. 1, no. 1, pp. 92–97, Feb. 2022.
doi:10.35842/icostec.v1i1.16
[19] G. Jozanikohan, G. H. Norouzi, F. Sahabi, H. Memarian, and B. Moshiri, “The application of
Multilayer Perceptron Neural Network in volume of Clay Estimation: Case study of shurijeh gas
reservoir, northeastern Iran,” Journal of Natural Gas Science and Engineering, vol. 22, pp. 119–131,
Jan. 2015. doi:10.1016/j.jngse.2014.11.022
[20] B. Gasperini et al., “Is the fast-track process efficient and safe for older adults admitted to the
emergency department?,” BMC Geriatrics, vol. 20, no. 1, Apr. 2020. doi:10.1186/s12877-020-
01536-5
[21] S_M Hosseini_Shokouh, K. Mohammadi and M. Yaghoubi, "Optimization of Service Process in
Emergency Department Using Discrete Event Simulation and Machine Learning Algorithm," Arch
Acad Emerg Med, vol. 10, no. 1, 2022. https://doi.org/10.22037/aaem.v10i1.1545
[22] C. Pope, G. McKenna, J. Turnbull, J. Prichard, and A. Rogers, “Navigating and making sense of
urgent and emergency care processes and provision,” Health Expectations, vol. 22, no. 3, pp. 435–
443, Jan. 2019. doi:10.1111/hex.12866
[23] D. A. Ghazali et al., “Profile and motivation of patients consulting in emergency departments while
not requiring such a level of care,” International Journal of Environmental Research and Public
Health, vol. 16, no. 22, p. 4431, Nov. 2019. doi:10.3390/ijerph16224431
[24] T. Morris, S. M. Mason, C. Moulton, and C. O’Keeffe, “Calculating the proportion of avoidable
attendances at UK emergency departments: Analysis of the royal college of emergency medicine’s
Sentinel Site Survey Data,” Emergency Medicine Journal, vol. 35, no. 2, pp. 114–119, Oct. 2017.
doi:10.1136/emermed-2017-206846
[25] https://github.com/karajeh-ala/Classifying-Emergency-Patients

AUTHORS
Ala’ Karajeh is a biomedical engineer and a recent master’s graduate from the
University of Manitoba. He worked under the supervision of Dr. Rasit Eskicioglu at
the Internet of Things lab to explore his interest in biomedical engineering and gait
analysis. His research interests include health monitoring and analytics and utilizing
cutting-edge technologies to enhance healthcare and medicine. He also examined
mobile application development, wearables and ECG signal analysis.

Dr. Rasit Eskicioglu is an Associate Professor at the computer science department,

University of Manitoba, Canada. He is also a research affiliate at the Center on Aging.
His research interests are experimental systems, particularly computer systems,
systems software, operating systems, distributed, cluster, and grid computing, high-
speed network interconnects, mobile networks, and pervasive computing. Most
recently, he is looking at wireless sensor networks (WSNs) and their applications for
solving real-world problems, such as indoor localization, monitoring, and tracking, and
the Internet of Things(IoT).

64
International Journal of Artificial Intelligence and Applications (IJAIA), Vol.15, No.3, May 2024

65