Predictive Modeling for COVID-19 Emergency Department

Stay using Machine Learning

Dr. M. Dhanalakshmi, professor of IT, G. Anjali, MCA Student,

Department of IT, Jawaharlal Nehru technological University, India

ABSTRACT: The COVID-19 pandemic patients. The predictive model could serve as a
has significantly increased emergency decision-making tool, aiding in resource
department (ED) stays for patients in the United planning for EDs and hospitals. Moreover, it
States. To address this issue, a study aimed to offers patients estimations of their ED LOS,
create a reliable model predicting the length of enhancing their understanding and potentially
stay (LOS) for COVID-19 patients in the ED improving their experience. In summary, this
and identify factors influencing meeting the '4- study's model effectively predicts ED LOS for
hour target.' Data from diverse urban hospitals in COVID-19 patients, enabling better resource
Detroit, collected from March 16 to December allocation and informed decision-making in
29, 2020, informed this research. Using data managing ED stays during the pandemic,
processing, four machine learning models potentially improving patient care and
(logistic regression, gradient boosting, decision hospital efficiency.
tree, and random forest) were trained to forecast
Keywords – COVID-19, length of stay (LOS), 4-
whether COVID-19 patients' ED stays would
hour target, emergency department (ED),
surpass 4 hours. The study involved 3,301
machine learning.
patients with 16 clinical factors. The gradient
boosting (GB) model outperformed others, [Link]
achieving 85% accuracy and an F1-score of 0.88
in predicting LOS within the test data, The coronavirus (COVID-19) pandemic has put

surpassing the logistic regression baseline, a burden on healthcare systems throughout the

decision tree, and random forest models. Further globe by increasing treatment complexity, the

data splitting did not notably improve accuracy. requirement for medical staff and patient safety,

This investigation identified critical factors, and an increase in patients suspected or infected

including patient demographics, existing health with the severe acute respiratory syndrome

conditions, and operational ED data, as coronavirus (SARS-CoV2). The inflow of

predictors of extended stays for COVID-19 infected COVID-19 patients at hospital

emergency rooms (EDs) has put current services [Link] REVIEW
under pressure. As a consequence of the
pandemic, multiple health-care institutions in the Effect of emergency department crowding on

United States have reported increased workload outcomes of admitted patients:

and spikes in patient volumes, resulting in ED

Crowding in the emergency department (ED) is
congestion, which worsens patient outcomes and
a common health-care delivery issue that may
places extra burden on medical personnel. The
have a negative impact on the outcomes of
creation of lineups in different sectors of the
patients who need hospitalisation. We
health system as a consequence of demand
investigate the relationship between ED
surpassing capacity is a significant feature of
congestion and subsequent outcomes in a wide
crowding. These queue formations are often
group of hospitalised patients. Methods We
associated with increased average ED lengths of
conducted a retrospective cohort study of
stay (LOS) . A longer ED stay is related with
patients admitted via the emergency departments
increased morbidity and death. Numerous health
of nonfederal, acute care hospitals in California
systems have established time-based standards,
in 2007. Inpatient mortality was the main
requiring patients to leave the ED within four
outcome. Hospital duration of stay and expenses
hours of their admission (i.e., the "four-hour
were secondary outcomes. The proxy measure of
target"). However, because to the continuing
ambulance diversion hours on the day of
epidemic, this 4-hour goal for COVID-19
admission was used to determine ED congestion.
patients has been difficult to meet, resulting in
To account for hospital-level confounders of
congestion, operational inefficiencies, and
ambulance diversion, we characterised high ED
increased consumption of hospital resources.
congestion as days that were within the top
quartile of diversion hours for a single
institution. Demographics, temporal factors,
patient comorbidities, main diagnosis, and
hospital fixed effects were all adjusted for in
hierarchical regression models. We employed
bootstrap sampling to quantify the extra
outcomes caused by ED crowding. Results We
looked at 995,379 ER visits that resulted in
admission to 187 hospitals. Patients treated on
days with high ED congestion had a 5% higher
risk of inpatient mortality (95% CI 2% to 8%), a
Fig.1: Example figure
0.8% longer hospital duration of stay (95% CI
0.5% to 1%), and a 1% increase in expenses per of comparable patients in the emergency
admission (95% CI 0.7% to 2%). Excess department on the same shift. The adjusted odds
outcomes due to high ED congestion included ratio (95% confidence interval) for death and
300 inpatient deaths (95% CI 200 to 500 admission in high acuity patients was 1.79 (1.24
inpatient deaths), 6,200 hospital days (95% CI to 2.59) for death and 1.95 (1.79 to 2.13) for
2,800 to 8,900 hospital days), and $17 million in admission in low acuity patients was 1.71 (1.25
expenses (95% CI $11 to $23 million). to 2.35) for death and 1.66 (1.56 to 1.76) for
Conclusion High ED congestion was linked to admission in high acuity patients. Leaving
higher inpatient mortality as well as minor without being seen was not connected with an
increases in length of stay and expenses for increase in adverse events at the patient level or
admitted patients. in hospital yearly rates. Conclusions Presenting
to an emergency department during shifts with
Association between waiting times and short
longer waiting times, as represented in a longer
term mortality and hospital admission after
mean duration of stay, is related with a higher
departure from emergency department:
risk of mortality and hospitalisation among
Population based cohort study from Ontario,
patients who are healthy enough to leave the
Canada
department in the immediate term. Patients who

To see whether patients who are not admitted to depart without being seen are not more likely to

the hospital after visiting an emergency have short-term harmful outcomes.

department during long-waiting shifts are at risk

Measures of crowding in the emergency
for adverse occurrences. Design Using health
department: A systematic review
administrative records, a population-based
retrospective cohort research was conducted. Despite agreement on the conceptual
Setting Emergency rooms with a high number of underpinning of crowding and rising research on
patients in Ontario, Canada, fiscal years 2003-7. crowding determinants and effects, there is no
Participants All non-admitted emergency criteria or measure of crowding. The goal was to
department patients (seen and discharged; left provide a thorough evaluation of crowding
without being seen). Outcome metrics The risk metrics and compare their conceptual base and
of adverse events ( hospitalisation or death validity. Methods: A systematic, thorough
within seven days) was adjusted for relevant evaluation of four medical and health care
patient, shift, and hospital factors. Results 13 citation databases was conducted to uncover
934 542 patients were seen and dismissed, while research linked to emergency department
617 011 were not seen. The risk of adverse congestion (ED). Publications that "explain the
outcomes rose with the average duration of stay theory, development, implementation,
assessment, or any other component of a validation across many locations and explain
'crowding measurement/definition' instrument which possibilities emerge as the metrics of
(qualitative or quantitative)" were eligible for choice in this "busy" area of measurements.
inclusion. A "measurement/definition" tool is
Systematic review of emergency department
anything that lends a numerical value to the
crowding: Causes, effects, and solutions
occurrence of congestion in the emergency
department. The following information was Crowding in emergency departments (EDs) is a
gathered from studies that met the inclusion global concern that may have an impact on
criteria: research design, objective, crowding health-care quality and access. We did a
measure, and evidence of validity. All thorough PubMed search to find studies that (1)
measurements were classified into five kinds investigated the causes, impacts, or solutions of
(clinician opinion, input factors, throughput ED crowding; (2) provided data collection and
factors, output factors, and multidimensional analytic technique; (3) took place in a general
scales). All metrics were then indexed to six ED context; and (4) focused on daily crowding.
validation criteria (clinician opinion, ambulance The relevant publications were determined by
diversion, time to care, forecasts or projections agreement by two independent reviewers. Each
of future congestion, and other). The databases study's methodology was graded using a 5-level
found 2,660 documents; 46 of these papers quality rating measure. The reviewers found 93
satisfied inclusion criteria, were original publications that met the inclusion criteria from
research investigations, and were abstracted by 4,271 abstracts and 188 full-text articles. A total
reviewers. A total of 71 distinct crowding of 33 papers investigated the causes, 27 articles
measurements were discovered. Clinician investigated the impacts, and 40 articles
opinion was the least generally utilised form of investigated the remedies to ED crowding.
crowding metric, whereas numerical counts Nonurgent visits, "frequent-flyer" patients,
(number or percentage) of patients and process influenza season, insufficient staffing, inpatient
durations connected with patient care were the boarding, and hospital bed shortages were all
most regularly employed. Many of the metrics often investigated causes of crowding. Patient
demonstrated a moderate to strong association death, transit delays, treatment delays,
with the validation criteria. Conclusions: Time ambulance diversion, patient elopement, and
intervals and patient counts seem to be the most financial impact were all often investigated
promising instruments for monitoring flow and impacts of crowding. Additional people,
nonflow (i.e., crowding). Standardized observation units, hospital bed access, nonurgent
definitions of time intervals (flow) and referrals, ambulance diversion, destination
numerical counts (nonflow) will help with control, crowding measures, and queuing theory
were all often investigated crowding solutions. severity score, abbreviated injury score (AIS)
The findings demonstrated the complicated, chest, and AIS head. Conditional logistic
multidimensional nature of the ED crowding regression was used to compare ED LOS
issue. Additional high-quality research may between the two groups. We found 509 blunt
contribute significantly to a better understanding trauma patients who needed to be intubated right
and alleviation of the everyday situation. This away. Thirty-three of these patients developed
organised literature review may aid in pneumonia and could be compared with
identifying future paths for the crowding equivalent controls. The case patients had a
research agenda. mean age of 44.6 (24.3), an injury severity score
of 32.7 (9.5), a chest AIS of 1.5 (1.6), and a head
Emergency department length of stay: A
AIS of 4.4 (1.2). The cases' ED LOS was
major risk factor for pneumonia in intubated
substantially longer than the controls' (281.3
blunt trauma patients
minutes vs. 214.0 minutes, p 0.05). Each hour

Pneumonia is a major cause of morbidity and raised the likelihood of acquiring pneumonia by

death in intubated patients. Pneumonia almost 20%. Conclusions: Increased ED LOS is

prevention measures have shown to be efficient an independent risk factor for pneumonia in

in the critical care unit and are well-liked, cost- blunt trauma patients who are intubated

effective, and effective. In the prehospital or urgently. Ventilator-associated pneumonia

emergency department (ED), trauma victims are therapies, which have been shown to be effective

often intubated on the spot. Hospital in the intensive care unit, should be started early

overcrowding has resulted in increased ED in the hospital course, and efforts should be

length of stay throughout the country (LOS). We made to reduce hospital congestion and ED

wanted to look at the link between lengthy ED LOS.

wait times and pneumonia rates. Methods: This

3. METHODOLOGY
was a 2-year retrospective case-control study of
pneumonia risk in blunt trauma patients Previous research on parameters linked with ED
admitted to an urban Level I trauma hospital and LOS done before to the COVID-19 pandemic
intubated immediately. Demographic and included models such as multiple linear
clinical data were obtained from the trauma regression, logistic regression, decision trees,
registry. All patients who were intubated and accelerated failure time models. Machine
prehospital or in the emergency department and learning algorithms may take into account a
acquired pneumonia were considered cases. greater number of characteristics and
There was a set of matched controls who did not permutations (e.g., patient records and hospital
get pneumonia and had the same age, injury information), which has the ability to provide a
better understanding of complicated issues and Fig.2: System architecture
uncover factors that predict COVID-19 ED
MODULES:
patients' LOS. To our knowledge, no research
has integrated these data (patient and ED To carry out the aforementioned project, we
operational data) to predict the LOS of COVID- created the modules listed below.
19 ED patients.
 Data exploration: we will put data into
Disadvantages: the system using this module.

1. No research has integrated these data (patient  Processing: we will read data for
and ED operational data) to predict the LOS of processing using this module.
COVID-19 ED patients.
 Using this module, data will be
In this work, we used four machine learning separated into train and test groups.
approaches, namely logistic regression, gradient
 Model generation: Building the model -
boosting, decision trees, and the random forest
Gradient Boosting, Random Forest,
algorithm, to construct a model that effectively
Decision Tree, Logistic Regression,
predicted the ED LOS of COVID-19 patients
XGBoost, and Voting Classifier.
across multiple data processing stages.
Calculated algorithm accuracy.
Advantages:
 User signup and login: Using this
1. improving ED and hospital resource planning module will result in registration and
and informing patients about improved ED LOS login.
projections.
 User input: Using this module will result
in predicted input.

 Prediction: final predicted shown

4. IMPLEMENTATION

ALGORITHMS:

Random Forest: A Supervised Machine

Learning Algorithm that is commonly utilised in
Classification and Regression applications. It
constructs decision trees from several samples
and uses their majority vote for classification regression and classification applications, among
and average for regression. other things. It returns a prediction model in the
form of an ensemble of weak prediction models,
Decision Tree: Decision trees use numerous
usually decision trees.
methods to determine whether or not to divide a
node into two or more sub-nodes. The
development of sub-nodes promotes the
5. EXPERIMENTAL RESULTS:
homogeneity of the sub-nodes that arise. In other
words, the purity of the node rises in relation to
the target variable.

Logistic Regression: Logistic regression is a

statistical analytic approach that uses past
observations of a data set to predict a binary
result, such as yes or no. A logistic regression Fig.3: Home screen
model forecasts a dependent variable by
examining the connection between one or more
existing independent variables.

Voting classifier: A voting classifier is a

machine learning estimator that trains numerous
base models or estimators and predicts based on
the results of each base estimator. Aggregating
criteria may be coupled voting decisions for
each estimator output.

XGBoost: Extreme Gradient Boosting Fig.4: User registration

(XGBoost) is a scalable, distributed gradient-

boosted decision tree (GBDT) machine learning
framework. It is the top machine learning
package for regression, classification, and
ranking tasks, and it supports parallel tree
boosting.

Gradient boosting: Gradient boosting is a

machine learning approach that is utilised in
Fig.5: user login patients throughout hospitalisation. The research
revealed significant characteristics linked with
extended stays in COVID-19 patients based on a
mix of patient demographics, comorbidities, and
ED operational data. We used these
characteristics to train four prediction models to
predict COVID-19 patients' ED LOS. The model
and findings of this research, with additional
validation, might serve as an effective decision-
Fig.6: Main screen support tool to enhance healthcare
delivery/resource planning and assist clinicians
in developing appropriate treatments to address
patient outcomes (e.g., reducing prolonged
LOS). Although the models were built using
locally obtained data and clinical information
from Henry Ford Hospital, they may be
retrained and updated to predict COVID-19
patient LOS in other EDs.

Fig.7: User input REFERENCES

[1] E. Walters, S. Najmabadi, and E. Platoff,

‘‘Texas hospitals are running out of drugs, beds,
ventilators and even staff,’’ Texas Tribune,
Austin, TX, USA, Tech. Rep., 2020.

[2] B. C. Sun, R. Y. Hsia, R. E. Weiss, D.

Zingmond, L.-J. Liang, W. Han, H. McCreath,
and S. M. Asch, ‘‘Effect of emergency
department crowding on outcomes of admitted
Fig.8: Prediction result
patients,’’ Ann. Emergency Med., vol. 61, no. 6,
[Link] pp. 605–611, 2013. [Online]. Available:
[Link]
Finally, we present some of the medical and
i/S019606441201699X
emergency department features of COVID-19
[3] A. Guttmann, M. J. Schull, M. J. Vermeulen, cJRyns33RP1E4G5NmD73cFgjAlDaBsN4NHe
and T. A. Stukel, ‘‘Association between waiting q
times and short term mortality and hospital
[7] B. G. Carr, A. J. Kaye, D. J. Wiebe, V. H.
admission after departure from emergency
Gracias, C. W. Schwab, and P. M. Reilly,
department: Population based cohort study from
‘‘Emergency department length of stay: A major
Ontario, Canada,’’ Brit. Med. J., vol. 342, p.
risk factor for pneumonia in intubated blunt
d2983, Jun. 2011.
trauma patients,’’ J. Trauma: Injury, Infection
[4] U. Hwang, M. L. McCarthy, D. Aronsky, B. Crit. Care, vol. 63, no. 1, pp. 9–12, 2007.
Asplin, P. W. Crane, C. K. Craven, S. K.
[8] D. Liew, D. Liew, and M. P. Kennedy,
Epstein, C. Fee, D. A. Handel, J. M. Pines, N. K.
‘‘Emergency department length of stay
Rathlev, R. W. Schafermeyer, F. L. Zwemer, Jr.,
independently predicts excess inpatient length of
and S. L. Bernstein, ‘‘Measures of crowding in
stay,’’ Med. J. Aust., vol. 179, no. 10, pp. 524–
the emergency department: A systematic
526, Nov. 2003.
review,’’ Academic Emergency Med., vol. 18,
no. 5, pp. 527–538, May 2011. [9] C. Morley, M. Unwin, G. M. Peterson, J.
Stankovich, and L. Kinsman, ‘‘Emergency
[5] N. R. Hoot and D. Aronsky, ‘‘Systematic
department crowding: A systematic review of
review of emergency department crowding:
causes, consequences and solutions,’’ PLoS
Causes, effects, and solutions,’’ Ann. Emerg.
ONE, vol. 13, no. 8, Aug. 2018, Art. no.
Med., vol. 52, no. 2, pp. 126–136, 2008, doi:
e0203316.
10.1016/[Link].2008.03.014.

[10] N. Bobrovitz, D. S. Lasserson, and A. D.

[6] S. Clair, A. Staib, S. Khanna, N. M. Good, J.
M. Briggs, ‘‘Who breaches the four-hour
Boyle, R. Cattell, L. Heiniger, B. R. Griffin, A.
emergency department wait time target? A
J. Bell, J. Lind, and I. A. Scott, ‘‘The national
retrospective analysis of 374,000 emergency
emergency access target (NEAT) and the 4-hour
department attendances between 2008 and 2013
rule: Time to review the target,’’ Med. J. Aust.,
at a type 1 emergency department in England,’’
vol. 204, no. 9, p. 354, 2016. [Online].
BMC Emergency Med., vol. 17, no. 1, p. 32,
Available: https://
Dec. 2017.
[Link]/doi/pdf/10.5694/mja15.
01177?casa_token=ijKnv
DmdfIIAAAAA:8dfpe4v4DN0Yo6d6yB–
7f5CIdrUiFV5BZ6yq61odPXx