Digital transformation to mitigate emergency situations: increasing opioid overdose survival rates through explainable artificial intelligence

Abstract

Purpose
The global health crisis represents an unprecedented opportunity for the development of artificial intelligence (AI) solutions. This paper aims to integrate explainable AI into the decision-making process in emergency scenarios to help mitigate the high levels of complexity and uncertainty associated with these situations. An AI solution is designed to extract insights into opioid overdose (OD) that can help government agencies to improve their medical emergency response and reduce opioid-related deaths.

Design/methodology/approach
This paper employs the design science research paradigm as an overarching framework. Open-access digital data and AI, two essential components within the digital transformation domain, are used to accurately predict OD survival rates.

Findings
The proposed AI solution has two primary implications for the advancement of informed emergency management. Results show that it can help not only local agencies plan their resources for timely response to OD incidents, thus improving survival rates, but also governments to identify geographical areas with lower survival rates and their primary contributing factor; hence, they can plan and allocate long-term resources to increase survival rates and help in developing effective emergency-related policies.

Originality/value
This paper illustrates that digital transformation, particularly open-access digital data and AI, can improve the emergency management framework (EMF). It also demonstrates that the AI models developed in this study can identify opioid OD trends and determine the significant factors improving survival rates.

Full text

Rapid #: -18159518
CROSS REF ID: 1048784
LENDER: IRU :: Main Library
BORROWER: WTU :: Olin Library
TYPE: Article CC:CCG
JOURNAL TITLE: Industrial management + data systems
USER JOURNAL TITLE: Industrial Management & Data Systems
ARTICLE TITLE: Digital transformation to mitigate emergency situations: increasing opioid overdose survival rates
through explainable artificial intelligence
ARTICLE AUTHOR: Johnson, Marina
VOLUME: ahead-of-print
ISSUE: ahead-of-print
MONTH: 10
YEAR: 2021
PAGES: -
ISSN: 0263-5577
OCLC #:
Processed by RapidX: 10/20/2021 12:09:58 PM
This material may be protected by copyright law (Title 17 U.S. Code)

Digital transformation to mitigate
emergency situations: increasing
opioid overdose survival rates
through explainable
artificial intelligence
Marina Johnson and Abdullah Albizri
Feliciano School of Business, Montclair State University, Montclair, New Jersey, USA
Antoine Harfouche
Paris-Nanterre University, Nanterre, France, and
Salih Tutun
John M. Olin Business School, Washington University in St. Louis, St. Louis, Missouri, USA
Abstract
Purpose –The global health crisis represents an unprecedented opportunity for the development of artificial
intelligence (AI) solutions. This paper aims to integrate explainable AI into the decision-making process in
emergency scenarios to help mitigate the high levels of complexity and uncertainty associated with these
situations. An AI solution is designed to extract insights into opioid overdose (OD) that can help government
agencies to improve their medical emergency response and reduce opioid-related deaths.
Design/methodology/approach –This paper employs the design science research paradigm as an
overarching framework. Open-access digital data and AI, two essential components within the digital
transformation domain, are used to accurately predict OD survival rates.
Findings –The proposed AI solution has two primary implications for the advancement of informed
emergency management. Results show that it can help not only local agencies plan their resources for timely
response to OD incidents, thus improving survival rates, but also governments to identify geographical areas
with lower survival rates and their primary contributing factor; hence, they can plan and allocate long-term
resources to increase survival rates and help in developing effective emergency-related policies.
Originality/value –This paper illustrates that digital transformation, particularly open-access digital data
and AI, can improve the emergency management framework (EMF). It also demonstrates that the AI models
developed in this study can identify opioid OD trends and determine the significant factors improving
survival rates.
Keywords Digital transformation, Explainable artificial intelligence, Machine learning, Emergency
management framework, Opioid OD, Survival prediction
Paper type Research paper
1. Introduction
Digital transformation refers to the integration of digital technology into all areas of a
business or an organization, fundamentally changing their processes and operations (Vial,
2019). Digital transformation, artificial intelligence (AI) and machine learning (ML)
revolutionize various industry domains (Kapletia et al., 2019;Roscoe et al., 2019;Wamba
and Queiroz, 2020). Such disruptive technologies are garnering significant attention due to
their many advantages, including maturity speed, limitless possibilities and power of
transforming organizations (Bohr and Memarzadeh, 2020;Delmolino and Whitehouse, 2018).
Following the private sector, governments at all levels have started employing digital
transformation, AI and ML in order to deliver services and programs more productively,
transparently and cost-effectively (Luna-Reyes and Gil-Garcia, 2014). Many government
agencies collect and store data as part of their digital transformation efforts and encourage
Artificial
intelligence for
emergency
The current issue and full text archive of this journal is available on Emerald Insight at:
https://www.emerald.com/insight/0263-5577.htm
Received 21 April 2021
Revised 14 September 2021
Accepted 25 September 2021
Industrial Management & Data
Systems
© Emerald Publishing Limited
0263-5577
DOI 10.1108/IMDS-04-2021-0248

researchers to extract insights and create applications for data-driven decision-making
through AI (Luna-Reyes and Gil-Garcia, 2014).
The outbreak of the COVID-19 pandemic has given rise to a dramatic increase in the opioid
crisis, a disruptive situation that has significantly increased the number of weekly visits to
hospital emergency departments. Compared to prepandemic rates, opioid overdose (OD) has
increased by 45% in this context (Nissen, 2021). In pre-COVID times, the opioid epidemic was
already ranked by government agencies and health organizations as one of the worst public
health crises in a generation with far-reaching social and economic effects (Centers for
Medicare and Medicaid Services, 2018). Opioids are a class of prescription drugs used for pain
relievers, including not only oxycodone, hydrocodone, fentanyl and tramadol, but also the
illegal drug heroin (NIH, 2020). Opioid addiction is defined by a strong and compulsive urge to
use opioid drugs due to their ability to produce feelings of pleasure, even when they are no
longer required medically (MedlinePlus –US National Library of Medicine, 2020). Opioid OD
refers to taking high doses of opioids (WHO, 2020). Opioid OD is signaled by pinpoint pupils,
unconsciousness and breathing troubles and can cause death when the brain function of
regulating breathing is seriously damaged (WHO, 2020).
The death toll of opioid OD in the USA alone stood around 450,000 between 1999 and 2018
(CDC, 2020). Moreover, opioid OD deaths were four-folds more in 2018 vis-
a-vis 1999,
indicating an accelerating trend of the opioid epidemic (CDC, 2020). Globally, 115,000
individuals died of opioid OD in 2017 only, while the number of nonfatal OD cases was several
times higher than the fatal OD cases (WHO, 2020). The annual economic burden and societal
costs (e.g. healthcare and loss in wages) of the opioid epidemic are estimated to be as high as
$1.02 trillion in 2017 in the USA alone (Florence et al., 2021). While guidelines and
recommendations to address the opioid epidemic are available, the gap between suggested
courses of action and reality is significant (WHO, 2020). Therefore, government agencies,
private organizations and emergency response teams need to mitigate, plan and direct their
resources to reduce the impacts of the opioid epidemic.
There have been growing investments in AI interventions to combat the opioid-driven OD
epidemic plaguing the entire world (Ti et al., 2021). We aim to examine the use of digital
transformation –particularly open-source digital government data –and AI in helping
government agencies and organizations to address opioid OD incidents, which account for an
emergency crisis. More specifically, this study’s primary research objective is to investigate
the role of AI in emergency mitigation and preparedness efforts concerning opioid OD and
provide government agencies with the adequate relevant insights. Additionally, we address
the following research questions throughout this study: (1) What are the important factors
increasing the survival rates of victims after an OD incident? (2) How do these factors
contribute to OD survival rates?
To attain the research objective and answer the aforementioned research questions, we
construct an information technology(IT) artifact using thedesign science research paradigm as
an overarching framework (Abbasi et al., 2012;Hevner et al.,2004). Based on the design science
research paradigm, this paper presents an IT artifact (i.e. an AI-based solution) utilizing state-
of-the-art AI algorithms. We employ the emergency management framework (EMF)
(McLoughlin, 1985) as the kernel theory guiding the design of the IT artifact. To build this
AI-based solution, we curate a dataset that includes the OD incidents of the past two years, from
the open data portal of the US government. We then create an explainable AI framework that is
consistently capable of predicting OD survival rates and identifying important variables. This
research has two primary implications in medical emergency situations: (1) it can help local
agencies plan their resources for a timely response to OD incidents, thus improving survival
rates. (2) It can be harnessed by state and federal governments to identify geographical areas
with lower survival rates and their primary contributing variables while planning and
allocating long-term resources to increase survival rates.
IMDS

The remainder of the paper includes a literature review in Section 2 and the description of
the theoretical development in Section 3.Section 4 outlines the proposed methodology, while
Section 5 presents the results and provides insights into the factors affecting OD survival
rates. The study’s implications are described in Section 6, while Section 7 serves as the
concluding remarks and future research directions.
2. Literature review
In this section, we conduct a literature review and examine published articles on the use of
digital transformation, AI, ML and emergency situations. We use Scopus to obtain the
relevant research because it is regarded as the largest abstract and citation database of
peer-reviewed literature. We employ two different search queries to capture all relevant and
important studies. In the first query, we aim to extract the articles studying digital
transformation, AI and ML in the context of emergency situations by using various keywords,
such as “Digital Transformation,”“Artificial Intelligence,”“Machine Learning,”and
“Emergency Situations.”We then limit the search results to the articles published in
business-related journals since 2010 by using Australian Business Deans Council (ABDC) and
Association of Business School (ABS) journal rankings. In the second query, we extract the
research studies exploring the use of digital transformation, AI and ML in the context of opioid
OD. We use various keywords (e.g. “Artificial Intelligence,”“Machine Learning”and “OD”).
Several studies utilize digital transformation, along with AI, for effective emergency
management (Cavdur and Sebatli, 2019;Chaudhuri and Bose, 2020;Fertier et al., 2020;Hayes
and Kelly, 2018;Lee and Lee, 2021;Pekar et al., 2020;Zhu et al., 2021). Some of these studies
focus on resource allocation to mitigate the impact of emergency situations. For example,
Chaudhuri and Bose (2020) apply deep learning to images of earthquake-hit regions to
identify survivors. They propose that their research can be utilized to better allocate the
search and rescue resources. Zhang et al. (2016) develop a resource allocation model for
emergency rescue using two-stage mixed-integer programming to optimize resource
allocation for rescue service within the rescue time horizon. Zhang et al. (2012) employ
linear programming and network optimization to allocate resources to emergency incident
locations. They take into account a potential secondary emergency situation and determine
the priority of preference regarding how the resources are allocated. Several studies focus on
minimizing the response time to emergency situations. For instance, Cavdur and Sebatli
(2019) develop a decision support tool to allocate temporary disaster response facilities using
ML to reduce the response times. Erdoǧan et al. (2010) employ the tabu search algorithm to
determine the best emergency response team locations to minimize the incident response
times. Various studies use AI, Big Data, optimization and simulation to create emergency
response plans and evaluate their effectiveness. For example, Hayes and Kelly (2018) process
data from social media to create action plans for natural disaster response. De Maio et al.
(2011) propose a framework utilizing semantic web technologies and soft computing to devise
an emergency response plan. Tinguaro Rodr

ıguez et al. (2010) design a decision support
system for emergency response teams using a two-level knowledge-based methodology. Gul
et al. (2020) design a hybrid framework to evaluate the preparedness of emergency
departments by predicting demand using artificial neural networks (ANNs) and assess the
performance of emergency departments via discrete event simulation. Tang and Shen (2015)
develop an emergency response plan through AI and empirically test its effectiveness within
an emergency evacuation context after a typhoon incident.
With regard to the opioid OD literature, recent research has shown interest in the use of
digital transformation, particularly AI and ML to tackle the issue. Lo-Ciganic et al. (2021)
improve the accuracy of ML algorithms in predicting opioid OD risk and capture the social
determinants of OD risk by integrating data from human services and criminal justice health
Artificial
intelligence for
emergency

claims. Yao et al. (2020) extract suicidal posts from among opioid users on Reddit and classify
OD incidents as intentional or unintentional using ANNs, with a view to better understanding
the rationale for such users’behavior, providing new insights into the attitude of those
involved in the opioid epidemic. Boslett et al. (2020) use logistic regression and random forests
(RFs) to predict opioid involvement in unclassified drug ODs and estimate the number of fatal
opioid ODs. Prieto et al. (2020) develop a natural language processing method combined with
various ML algorithms (e.g. RF, k-nearest neighbors, support vector machines [SVMs] and
L1-regularized logistic regression) to identify potential opioid misuse from paramedic
documentation. Ward et al. (2019) develop an ML method to classify death certificates as drug
ODs to provide faster drug OD mortality surveillance and inform public health responses to
the drug OD epidemic in near-real time instead of several weeks following events. Badger
et al. (2019) develop ML models (i.e. RFs and regularized logistic regression) for classifying the
severity of opioid OD events from clinical data. Lo-Ciganic et al. (2019) utilize RF, gradient
boosting machine and deep neural network (DNN) approaches to predict opioid OD risk
among medicare beneficiaries. Dong et al. (2019) build ML and deep learning models to predict
opioid OD of patients based on both the history of patients’electronic health records and New
York State claims data. Neill and Herlands (2018) apply ML approaches to 17 years of county-
aggregated data for monthly opioid OD deaths in the New York City metropolitan area to
detect and characterize emerging OD patterns (e.g. geographic, demographic and behavioral
patterns). The literature review has demonstrated that there is no study exploring the opioid
epidemic and predicting opioid OD survival rates based on an explainable AI approach,
particularly in the context of emergency situations. This study bridges this gap by
developing AI-based models to predict opioid survival rates and identify its contributing
factors.
3. Theoretical development
Digital transformation, AI and ML can facilitate and improve emergency response activities
during pandemics, man-made or natural disasters, and disruptions. Therefore, we employ the
EMF (McLoughlin, 1985) as the kernel theory to guide the design of the IT artifact and
integrate it with AI and ML to improve its effectiveness. The EMF comprised four stages,
namely mitigation, preparedness, response and recovery.
The mitigation phase of the EMF aims to determine the actions that should be taken
before emergency incidents to eliminate or reduce their impacts and risks (McLoughlin, 1985).
This phase is significant to identify effective prevention plans and design standards to reduce
the risk of loss of life. Although traditional methods of collecting and organizing expert
opinions (e.g. surveys) can identify emergency risk factors, they are expensive, time-
consuming and not very accurate (Bellaire et al., 2017). AI can revolutionize this first phase by
analyzing large volumes of data using AI and ML. Based on insights from analyses, decision-
makers can develop effective mitigation strategies (Gama et al., 2016), such as identifying
management priorities (Canon et al.,2019) and developing contingency plans (Dou
et al., 2014).
The preparedness phase comprises identifying future emergency incidents and sending
out early warnings (McLoughlin, 1985). Traditionally, human experts measure, analyze and
plan emergency actions to deal with specific situations. AI can be a cost-efficient alternative
tool to predict potential emergency situations (i.e. disasters) and estimate the damage and
prepare an action plan (Ko and Kwak, 2012). AI has the potential to exhaustively analyze and
simulate various response plans to emergency situations with different magnitudes and
impacts, thus identifying the best plan (Sun et al., 2020). For example, it can help determine the
best evacuation plan or devise a resource allocation program after an earthquake (Zheng and
Liu, 2019).
IMDS

The response phase requires decision-makers to act decisively within tight time frames
(Sun et al., 2020) with often incomplete information and too much data from which it is
difficult to extract relevant information (i.e. information overload) (Carver and Turoff, 2007).
AI can be applied to collect, consolidate and analyze data rapidly, check its relevance and
reliability, and help decision-makers synthesize the best options for action (Ramchurn et al.,
2016). In an emergency, AI is able to rapidly classify information from a vast number of calls
and share the urgent needs of the victims to relevant emergency services (Sun et al., 2020).
AI can help automatically recognize voices, identify keywords and rapidly process voice data
(Ramchurn et al., 2016) to improve the response.
The recovery phase requires a quick understanding of emergency complexity, the
identification of operational needs and recovery plans, and the achievement of rehabilitation
and reconstruction activities (Sun et al., 2020). Traditionally, the loss and repair costs are
usually estimated based on tabulating data from different sources, such as insurance claims
and postdisaster assessments (Kim et al., 2016). However, the lack of standardized methods
for collecting and analyzing data may lead to different estimates (Ladds et al., 2017). The
availability of big data and the rapid development of AI contributes to assessing the disaster-
induced impact, establishing postevent recovery plans and conducting recovery and renewal
activities (Jamali et al., 2019). AI can also help identify potential fraud (Dutta et al., 2017) and
fake news (Zubiaga et al., 2018). Table 1 demonstrates the role of digital transformation and
disruptive technologies in improving the EMF.
4. Methodology
We develop an AI-based framework to achieve the research objective and answer the
research questions. This three-stage framework is illustrated in Figure 1. While the first stage
creates a dataset from various open data sources (e.g. OpenDataPA), the second trains three
AI algorithms –RF, ANN and SVM –to predict the survival probabilities of OD’ed people.
Lens/Phase Mitigation Preparedness Response Recovery
Industry lens:
Federal Emergency
Management
Agency –FEMA
(McLoughlin, 1985)
Identify long-
term risk of
disruptions
diminish long-
term risk
Prepare and plan
response
capabilities and
simulate their
effectiveness
React immediately
to emergency
situations
Timely support to
vital-life and work
to resume regular-
life
Academic lens
(Petak, 1985)
Determine
potential risk and
design risk
reduction
initiative
Present response
plan, identify
critical resources
and train first
responders
React to reduce
damage and
decrease
intervention time
Provide immediate
support to
temporary
activities and
install processes to
return to normalcy
Activities by
traditional
technologies
(Reddick, 2011)
Prevent
emergency
situations and
reduce damage
Enhance readiness
of organizations
Provide emergency
assistance
immediately
Provide pre-
emergency well-
being
Activities that can
be done by digital
transformation
Identify impact
and risks of
emergency and
predict potential
risk factors (Dou
et al., 2014)
Develop optimal
response plan and
help train
responders
(Mendonça and
Wallace, 2007)
Collect, consolidate
and analyze data
rapidly. Check its
relevance and
reliability
(Ramchurn et al.,
2016)
Assess loss and
devise strategy to
return to normalcy
and help
implement
strategy
(Deryugina, 2017)
Table 1.
AI for emergency
management
Artificial
intelligence for
emergency

During model training, cross-validation (CV) is used to prevent overfitting, which refers to
obtaining superior performance in the training set and poor performance in the test set, as the
model captures the signal and the noise in the training set. After training the AI models, the
third stage employs Shapley Additive Explanations (SHAP) to obtain insights into the factors
affecting survival rates. Emergency responders and policymakers can use these insights to
address OD incidents and improve the response plan, thus increasing the survival rates.
4.1 Data description and processing
Many government agencies in the USA have adopted digital transformation and taken
numerous steps to increase transparency and accountability to their citizens and taxpayers
(Luna-Reyes and Gil-Garcia, 2014). Notably, these agencies adopted open data initiatives that
promote increased access to government data to stimulate innovation and demonstrate the
effectiveness of their policies and programs (The United States Open Data, 2018). Open data –
particularly open government data –remain primarily untapped though it is an excellent
resource for obtaining insights and resolving societal and economic problems
(TheGovLab, 2020).
In 2018, many states in the USA started sharing their data with the public and launched
their open data initiatives (The United States Open Data, 2018), such as Pennsylvania’s (PA)
centralized and official data portal (OpenDataPA), which includes thousands of datasets that
are easily accessible to the public and can be used and redistributed without any restrictions.
The primary goal of OpenDataPA is to engage citizens and include researchers to help
government agencies make data-driven decisions and develop policy solutions (The Office of
Governor Tom Wolf, 2016).
Utilizing OpenDataPA as a primary source, this study curates a dataset to realize the
research objective and answer the primary research questions. This arduous data acquisition
process is described in Figure 2. The opioid OD incidents –used as the primary dataset –are
obtained from a source containing information on OD responses that were provided by PA’s
first responders (e.g. primarily police officers, secondarily firefighters and emergency medical
services staff) between December 2018 and November 2020 [1]. This data source includes
22,423 OD incidents with 34 variables, such as the overdosing person’s age, the drug
regularly taken by the victim OD’ed, the county where the OD incident took place, incident
response time and whether the overdosing person survived.
At the county level, more variables from secondary data sources on OpenDataPA are
added to enrich the OD dataset and provide more insights to government agencies and
policymakers. First, the number of drug and alcohol treatment facilities within the counties is
calculated using a dataset from OpenDataPA [2] and merged into the primary dataset. This
variable represents the level of access to treatment in the county. Several other proxy
Figure 1.
AI-based framework
IMDS

variables are obtained from the US Census Bureau and OpenDataPA and appended to the
primary dataset. These variables define the counties’financial health and resources and
include the number of registered businesses [3], unemployment rate [4], income per capita [5]
and poverty rate [6]. Lastly, the county’s population in which the OD incidents took place and
the number of transportation vehicles [7] they provide are merged with the primary dataset to
indicate if the county is considered urban or rural. The dataset used in this study is described
in Table 2.
4.2 AI algorithms
We use RF, ANN and SVM algorithms to train AI models based on the preliminary analysis.
RF is a tree-based supervised AI algorithm comprising multiple decision trees (DTs)
represented by “if-then-else rules”(Breiman, 2001). DTs within an RF model consist of nodes,
branches and leaves. The nodes stand for the test cases (e.g. “if statements”), while the
branches indicate the result of the test cases (e.g. “rules”)(Breiman, 2002). On the other hand,
the leaves represent the predicted outcome and are connected to the nodes via branches.
During the DT algorithm, the variables within the training set with the most discriminative
power to classify the instances (e.g. survived, deceased) are iteratively selected. These
selected variables become the nodes, and their values used to split the training set are
determined using a metric named information gain, indicating the reduction in entropy (e.g.
uncertainty, impurity). The RF algorithm uses the DT algorithm and builds many
uncorrelated DTs by sampling observations with replacement from the training dataset
(Cutler et al., 2012). Then, these individual DTs are combined using a function, such as simple
averages or majority voting. Since RF uses multiple uncorrelated DTs by sampling the
training set, it tends to provide lower model variance and better accuracy rates, which is
considered very robust. Readers interested in obtaining more information regarding RF are
referred to Breiman (2001) and Cutler et al. (2012).
The ANN is a supervised AI algorithm with self-learning capabilities to generate accurate
predictions and can model highly nonlinear relationships between input and target variables
Figure 2.
Data collection process
Artificial
intelligence for
emergency

(James et al., 2013). ANNs are composed of input, hidden and output layers, each containing
various artificial neurons. The input layer neurons represent the input variables and receive
the input data in the training set, while the output layer neuron stands for the target variable.
The hidden layer is located between the input and output layers. The hidden layer is used to
nonlinearly transform and combine the existing variables, so as to create new features that
Variable name Description Summary
Day Day of OD incident 7 levels (e.g. Monday)
County name County of OD incident 67 levels (e.g. Bucks, York)
Gender Gender of OD’ing person 3 levels (i.e. M, F, unknown)
Age range Age range of OD’ing person 10 levels (e.g. 30–39)
Race Race of OD’ing person 5 levels (e.g. White, Black)
Ethnicity Ethnicity of OD’ing person 3 levels (e.g. Hispanic)
Susp OD drug Drug person OD’ed on 19 levels (e.g. heroin)
Response time Response time to OD incident 6 levels (e.g. <3 min, >3 min
and <5 min)
Third party admin Responders to OD incident 4 levels (e.g. EMS, FD)
Season Season of OD incident 4 levels (e.g. Summer)
Naloxone administered Whether responders administered
Naloxone
2 levels (i.e. yes, no)
Year Year OD occurred in 3 levels (i.e. 2018, 2019, 2020)
Month Month of OD incident 12 levels (e.g. January)
Month day Day of the month OD occurred Min: 1, Mean: 15.2, Max: 31
Weekend Whether OD occurred on a weekend 2 levels (i.e. yes, no)
Time block Time block of OD incident 4 levels (e.g. 00:00–03:59)
Dose count Dose of Naloxone administered Min: 0, Mean: 1.0, Max: 9
Dose unit Dose count of Naloxone administered Min: 0, Mean: 1.7, Max: 4
Time Approximate time of OD incident Min: 0, Mean: 12.4, Max: 24
Number of trans Number of public transportation
vehicles
Min: 2, Mean: 9.3, Max: 19
Labor force Number of employed and unemployed
people
Min: 1800, Mean: 251079, Max:
709600
Employed Number of employed people Min: 1700, Mean: 239494, Max:
670300
Unemployed Number of unemployed people Min: 100, Mean: 11573, Max: 39300
Unemployment rate Percent of people unemployed Min: 0.9, Mean: 4.8, Max: 8.1
Number of registered
business
Number of corporations Min: 373, Mean: 73713, Max:
291719
Children in poverty Number of children living in poverty Min: 23, Mean: 4678, Max: 38729
Percent children in poverty Percent of children living in poverty Min: 8.1, Mean: 18.6, Max: 36.6
Per capita income Median income per person Min: 14325, Mean: 27770, Max:
41251
Median household income Median income per household Min: 34796, Mean: 54099, Max:
84741
Median family income Median income per family Min: 43750, Mean: 66769, Max:
94592
Number of households Number of households Min: 2273, Mean: 199573, Max:
599736
Number of drug treatment
facility
Number of drug/alcohol treatment
facilities
Min: 1, Mean: 33, Max: 121
Population Population Min: 4447, Mean: 481308, Max:
1584060
High pop Whether county of OD incident is rural
or not
3 levels (i.e. rural, mid-rural, urban)
Survive Whether OD’ed person survived 2 levels (e.g. yes, no)
Table 2.
Summary of variables
used in this study
IMDS

are not readily available in the training set (Ramachandran et al., 2017). The neurons located
within different layers are interconnected via weights determining the connection’s strength
(Haykin, 2009). The ANN first processes the external training set from the input layer to the
output layer in order to acquire the predicted target values. This step is called feedforward.
Then, the errors between the predicted and actual target values and their partial derivative
are computed. Another step consists of backpropagating the partial derivatives of the errors
with respect to weights, with a view to adjusting the weights. The ANN learns the optimal
weights that minimize the difference between the predicted and actual target variables by
iteratively feedforwarding the training set and backpropagating the errors.
The SVM is a supervised learning algorithm mainly used for classification. The SVM’s
goal is to find a hyperplane that optimally divides the dataset into two classes (i.e. died vs.
survived), with the most significant gap possible using quadratic programming (Cortes and
Vapnik, 1995;Noble, 2006). It is important to note that the SVM can utilize various kernel
functions (e.g. radial) to classify datasets that are not linearly separable. We use the sigmoid
kernel for the SVM algorithm based on our preliminary analysis, allowing it to operate in high
dimensional space with high accuracy rates efficiently.
4.3 Performance measures
In this study, RF and ANN classifiers are used to predict the class probabilities (e.g. the
likelihood of an overdosing person to survive). Therefore, these probabilities need to be
binarized using a discrimination threshold to determine true positive (TP), true negative (TN),
false positive (FP) and false negative (FN) rates. In the context of this study, TP represents an
outcome in which the AI model correctly predicts the positive class –overdosing person to
survive, while TN indicates an outcome in which the AI model accurately predicts the
negative class –overdosing person to decease. FP and FN are outcomes where the model
incorrectly predicts the positive and negative classes, respectively.
Primary performance metrics of AI models used for classification include accuracy,
sensitivity, specificity and the area under the curve (AUC). Accuracy indicates the percentage
of correctly classified observations (TP and TN). Sensitivity provides information regarding
the AI model’s ability to identify true positive (TP) instances (e.g. survived). Specificity
determines the AI models’ability to detect true negative (TN) instances (e.g. deceased). The
AUC is obtained by first creating a receiver operating characteristic (ROC) curve and
computing the area underneath. The ROC plots sensitivity against 1- specificity by varying
the threshold values used to binarize the class probabilities, thus eliminating the need for
selecting a discrimination threshold.
4.4 Data balancing
The categorical target variable used in AI models may have a class imbalance problem,
referring to one category (i.e. level) dominating the other. The category with more instances is
called the majority class, while the category dominated by the majority class is called the
minority class (Johnson et al.,2021). AI algorithms tend to learn the patterns in the majority
class, while ignoring the information in the minority class if the training dataset is imbalanced.
As a result, the AI algorithms tend to label most instances in the test set with the majority class
label, yielding low performance measures. Hence, it is critical to balance the training set during
CV while building AI models. Generally, oversampling and undersampling techniques are used
to balance the imbalanced training set (Chawla et al.,2011;Fern
andez et al.,2018). Observations
from the majority class are randomly removed in undersampling, while new synthetic
observations are added to the minority class in oversampling.
In the context of this study, the number of OD’ed people who survived is much higher than
the number of OD’ed people who died. To address this issue, an oversampling technique
Artificial
intelligence for
emergency

named Synthetic Minority Oversampling Technique (SMOTE) is used. The primary reason
for choosing SMOTE over other undersampling methods is to avoid removing observations
with valuable information. In SMOTE, a random observation from the minority class is first
selected, and then the distance between this observation and other observations in the same
class is computed. Using the distance values, the knearest neighbors of this random
observation are determined. One of these kneighbors is randomly selected, and a straight line
between this neighbor and the observation is drawn. Then, a new observation is created at
random on this straight line. The steps to create new observations are iterated until the
number of observations in the minority class is the same as the majority class. Readers
interested in learning more about SMOTE are referred to the articles written by Chawla et al.
(2011) and Fern
andez et al. (2018).
4.5 Cross-validation (CV)
K-fold CV is used to validate the AI models and prevent overfitting (Johnson et al., 2020;
Simsek et al., 2021). A completely independent dataset is needed to test the performance of the
AI models after K-fold CV. Thus, the dataset is split into training and test sets before K-fold
CV. The training and test sets contain 70 and 30% of the observations in the dataset. The
70:30 split ensures that the training data have enough instances with different patterns to fit
robust AI models. Additionally, the 70:30 split guarantees enough observations in the test set
to measure the AI models’performance and generalizability on the unseen dataset (Johnson
et al., 2021).
In K-fold CV, the training set is randomly partitioned into ksubsets. One of these ksubsets
is used as a test set, while the remaining k–1 subsets are put together to form the training set
used for training the AI model. In other words, an AI model is run ktimes, and performance
measures (e.g. AUC) across all ktrials are recorded. The final performance of the CV process is
acquired by averaging the performance measures of the kfolds. It is critical to note that the
data imbalance problem is addressed during CV. If the combination of k–1 subsets used as a
training set is imbalanced, SMOTE is used to correct that imbalance.
The advantage of k-fold CV is that each observation in the dataset gets to be used as a test
case. Determining the right kvalue is critical for obtaining robust AI models. High kvalues
increase computational run time and reduce the number of unique iterations with different
observations, leading to high model bias, referring to obtaining an erroneous model due to
missing relevant relations in the dataset. Low kvalues form drastically different k1 training
sets, which may increase the model variance, referring to the sensitivity to small fluctuations
in the dataset. Therefore, the kvalue should be selected by taking into account the
bias-variance tradeoff as well as the computational run time. This study uses five-fold CV (i.e.
k55) because previous studies have demonstrated that k55 provides a well-balanced trade-
off among computational run time, model bias and variance (Kohavi, 1995;Olson and Delen,
2008). The AUC is chosen as the primary metric to evaluate AI models’performance because
it does not require a discrimination threshold.
4.6 Model explainability
SHAP was developed by Lundberg and Lee (2017) and is used to explain and interpret AI
models at the individual observation level. The SHAP algorithm computes the contribution of
an input variable to a particular prediction as follows (Aas et al., 2020;Ancona et al., 2019;
Lundberg and Lee, 2017): (1) It builds models using all the permutations of the remaining
input variables with and without this specific feature; (2) The marginal contributions of this
feature are computed by subtracting the predictions obtained from these models; (3) These
marginal contributions across all permutations are averaged to determine the feature’s
Shapley value. Hence, the Shapley value of a variable at the individual observation level
IMDS

identifies how much this feature contributes to this observation’s prediction (Lundberg and
Lee, 2017). To obtain the SHAP feature importance, the absolute Shapley values of a feature
for individual observations are averaged.
5. Results
5.1 AI models’results
Table 3 summarizes the performance measures of the AI models obtained from the test set.
Table 3 shows the AUC rates for the AI models indicating that the RF model outperforms the
ANN and SVM models. We think RF’s ability to handle outliers and prevent overfitting may
play a role in its superior performance on the test set. Additionally, the set of “if-then-else”
rules created by the RF model may be well-suited for predicting OD survivals. For example, a
30-year-old person who is administrated Naloxone after a heroin OD may have a higher
chance of surviving than a 50-year-old who is not given Naloxone after a fentanyl OD. The RF
algorithm can model such relations with a high degree of accuracy using “if-then-else”rules,
yielding better results.
Table 3 also summarizes the accuracy, sensitivity and specificity rates of the AI models.
The RF model produces the best accuracy with an accuracy rate of 83.22%, showing that the
RF model inaccurately predicts the outcome of approximately 16 out of 100 OD instances.
Table 3 indicates that the RF model also yields better sensitivity and specificity rates than the
ANN and SVM models. The sensitivity indicates the model’s ability to identify the OD’ed
people who survived, while specificity shows how capable the AI model is in determining the
OD’ed people who did not survive.
5.2 Model explainability
The results described above indicate that the RF model performs the best to predict if an
OD’ing person will survive, given the input variables. The ANN model is the second-best
model. Thus, these two models are selected to be further explained. The ANN model has a
black-box nature and is not interpretable, and the interpretation of the RF model is limited.
Thus, SHAP is applied to these two AI models as part of the posthoc analysis. SHAP aims to
explore each input variable’s contribution to the prediction and identifies the critical input
variables that play a role in improving OD survival rates. Particularly, understanding the
relationship between the input and the target variables through SHAP can help government
agencies create policies and design interventions that can increase the survival rates after OD
instances.
The SHAP algorithm computes the Shapley values of input variables at the individual
observation level and determines how they contribute to this observation’s prediction. The
force plots, as provided in Figure 3, are used to demonstrate the Shapley values. The Shapley
value of each input variable is a “force”that either increases or decreases the prediction of a
particular observation. The base value on the force plots shows the average of all the
predictions in the test case. Each Shapley value on the force plot is an arrow displayed in
either blue or red, pushing to decrease or increase the prediction, respectively.
Performance measures RF ANN SVM
Accuracy % 83.22 79.33 76.54
Sensitivity % 82.06 77.95 74.01
Specificity % 87.36 84.25 80.04
AUC % 84.71 81.10 78.57
Table 3.
Performance measures
of AI models on test set
Artificial
intelligence for
emergency

Figure 3 provides four force plots, two for each AI model. Two OD incidents –one with a
surviving OD’ed person and one with a deceased OD’ed person –are randomly selected from
the test set to plot these force plots. It is critical to note that the force plots can be created for
each observation in the test set. Figure 3a suggests that the average predicted value for the
test set (i.e. the base value) is 0.4001, while the predicted target value for this observation is
0.25, indicating a low chance for the OD’ed person to survive. It is observed from this force
plot that the person OD’ed on fentanyl, and the response time to this emergency incident was
recorded as unknown. These specific values (i.e. fentanyl and unknown response time)
decreased this OD’ed person’s chances to survive. The OD’ed person, as shown in Figure 3a,
was administered a dose of Naloxone, which increased his/her chances to survive. Figure 3b
shows that the OD incident was responded to within a minute, and the OD’ed person was
given two doses of Naloxone, which increased his/her chances to survive. These results
indicate that the rapid response to an OD emergency case and administering Naloxone, when
the drug is heroin or fentanyl, is crucial to increasing survival rates.
Figure 3c suggests that the unknown response time and fentanyl decreased the OD’ed
person’s chance to survive while administering a dose of Naloxone increased his/her chances
to survive. In Figure 3d, rapid response to the incident raised the OD’ed person’s chances to
survive. The labor force and employed variables may be correlated with other input variables
and indirectly impact the OD’ed person’s likelihood to survive. For example, areas with
higher labor force and employment rates may have the resources to respond to emergency
OD incidents faster.
Figure 4 provides a summary plot that combines the importance of variables with effects.
Each point on the plot indicates a Shapley value of a feature for a given observation. The
position on the x-axis is the Shapley value, and the position on the y-axis is determined based
on the SHAP variable importance, which is the average of absolute Shapley values across all
Figure 3.
SHAP force plots
IMDS

observations. The color represents the value for the input variables. The overlapping points
are stacked vertically; thus, the Shapley values’overall distribution for a given input variable
can be observed.
According to the SHAP summary plot in Figure 4a, fast response time to OD instances
drastically reduces the predicted probability of death. Administering Naloxone increases the
chance to survive. Additionally, certain drugs encoded with higher numbers, such as heroin
and fentanyl, are substantially more dangerous, which may cause an OD’ing person to die.
Low age values reduce the risk of death, while high age values increase the risk of death.
People OD’ing earlier in the day and responded by police officers have higher chances of
surviving than people OD’ing later in the day and responded by a third party. This may be
related to two factors: (1) the person may be taking the first dose of the drug earlier in the day,
so the drug is not built up in his/her system, (2) there may be more available resources (e.g.
first responders, vehicles, etc.) earlier in the day so the OD incidents can be responded much
faster, thus reducing the risk of death.
According to the SHAP summary plot in Figure 4b, it is observed that the low response
time to an OD incident increases the risk of death, while high values in the population variable
(e.g. urban areas) decrease the risk of death. In other words, the risk of death after an OD in
urban areas is less than the rural areas. This may be associated with the resources that urban
areas have in the community. High values in employment, labor force, median household
income and low unemployment values increase the survival rate after an OD incident. This
means that people who OD’ed in areas with high employment and median household income
rates have higher chances of survival. These observations indicate that people have higher
chances to survive after an OD incident in counties with better resources.
Table 4 provides the SHAP variable importance values for each AI model obtained by
averaging the absolute Shapley values across all observations. According to this table,
response time to OD incidents is the most crucial variable that increases survival probability.
This variable is followed by administering Naloxone and its dose. Additionally, OD’ing on
specific drugs, such as fentanyl and heroin, drastically reduces the survival rates. The age
range variable is also critical in determining if the OD’ing person survives. According to the
dataset, OD’ed people between 40 and 60 have a significantly lower chance of surviving than
people younger than 40. The time block of the OD incident affects the probability of survival.
It appears that lower time block values indicating earlier times in a day increase the
survival rates.
Additionally, certain counties, such as Juniata, Warren and Clarion, have much lower
survival rates than the other counties, including Mercer, Alleghany and Philadelphia. All of
these afore-mentioned counties are less populated than the latter, indicating that urban areas
Figure 4.
SHAP summary plot
Artificial
intelligence for
emergency

better cope with OD instances. Variables that identify if a county is an urban or rural area and
are used as a proxy to define its resources positively affect the survival rates. For example,
the variables related to the population employed, number of households, median household
income and number of registered businesses have a direct relationship with survival rates,
which means that increased levels of population, number of households and median
household income improve the survival rates.
6. Discussions
6.1 Summary of key findings
This study examines the OD incidents in the USA, particularly in PA, and develops an
AI-based solution to predict opioid OD survival rates. The proposed AI-based solution
contains three stages. The first stage creates a dataset from various open data sources to
realize the research objective. Thus, results show that properly transforming, cleaning and
normalizing the input data is a must. Processing the input data is the key to the entire process.
Not every input feature in a given dataset will be informative for predicting the output labels.
Including irrelevant features can lead to overfitting, which refers to obtaining superior
performance in the training set and poor performance in the test set. Our study shows how CV
can be used to prevent such a risk.
The second phase trains AI algorithms. Our study develops three useful AI models –the
RF, ANN SVM models –to predict the survival probabilities of victims after an opioid OD.
This refers to a serious of back-and-forth steps aimed at finding the optimal set of model
parameters that translate the features in the input data into accurate predictions of the labels.
After many evaluations, error identifications, corrections and tests, the model performance
reaches its optimal level, while the model errors touch their minimum. Results show that the
RF model satisfies the best the solution objectives, since it yields high accuracy rates of 84.71.
These results indicate that the model can reliably predict if an OD’ed victim will survive,
given the input variables.
The third phase utilizes SHAP to make the models more transparent and interpretable.
This phase makes it possible to generate insights into the significant factors affecting OD
survival rates. Moreover, it identifies how the characteristics of the geographic locations in
which opioid OD occurs impact the response to the OD incident and the survival probability,
due to the SHAP ability to show how much each feature contributes to the target variable
prediction. Furthermore, it shows how features can influence the prediction movement
towards the two labels.
Variable RF ANN Average Variable RF ANN Average
Response time 0.249 0.020 0.135 Time 0.012 0.001 0.006
Dose count 0.046 0.000 0.023 Employed 0.002 0.009 0.006
Naloxone
administered
0.043 0.002 0.022 Labor force 0.003 0.006 0.005
Susp OD drug 0.039 0.002 0.021 Number of drug treatment
facility
0.009 0.001 0.005
Third party admin 0.022 0.002 0.012 Race 0.009 0.000 0.005
Age range 0.023 0.001 0.012 Number of registered
business
0.008 0.001 0.004
Time block 0.019 0.001 0.01 Children under in poverty 0.005 0.001 0.003
County name 0.018 0.001 0.009 Median household income 0.003 0.003 0.003
Dose unit 0.015 0.001 0.008 Number of households 0.005 0.002 0.003
Population 0.003 0.012 0.008 Day 0.006 0.001 0.003
Table 4.
SHAP variable
importance –top 20
IMDS

6.2 Theoretical implications and contributions
Digital transformation capabilities and the availability of data and the rapid development of
ML applications offer an unprecedented opportunity to promote AI solutions. This paper
demonstrates the importance of integrating AI solutions into decision-making processes in
emergency scenarios to help mitigate the high levels of complexity and uncertainty
associated with these situations.
Our proposed AI solution covers different types of factors, contributing to the
advancement of informed emergency management at different phases. The two phases of
emergency mitigation and preparedness are the most complex because they cannot be easily
anticipated (Waugh and Streib, 2006).
However, as shown in the results section, the proposed AI solution can support emergency
mitigation management while making emergency organizations more resilient. Indeed, it can
help decision-makers in identifying potential risks based on geographical and socioeconomic
factors, predict possible impact, assess vulnerability, gain better situation awareness with
more confidence and help decision-makers to develop effective mitigation strategies and
emergency-related policies.
The proposed AI solution can also help in the preparedness phase when it comes to
identifying the upcoming emergencies in real time. Indeed, AI outperforms humans in terms
of data analysis speed and the volume of analyzable data that can be processed. As shown in
the results section, the proposed AI can help in proposing accurate predictions and
suggestions that may be needed by decision-makers to deploy resources and develop
emergency plans in an optimal way. It is a powerful and a cost-effective tool to support
decision-making in emergencies where speed is a matter of life and death.
The two phases of response and recovery are multifaceted processes, involving many
actors. As shown in the results section, the proposed AI solution can help these numerous
actors in quickly sharing the same understanding of the situation and, therefore, supporting
the emergency team in improving the efficiency of its response efforts.
As emergency recovery usually takes a long time, the proposed AI solution can be an
important module in quickly completing the damage assessment and supporting the
recovery in less time. Therefore, emergency organizations need to integrate AI tools with
their existing ICT to improve their performance in emergency recovery.
6.3 Practical implications and contributions
The center for disease control and prevention reports that 67,367 opioid OD deaths occurred
in the USA in 2018. Thus, there is a need for governments to use a solution similar to the one
proposed in this study to identify the primary variables increasing survival rates of OD
incidents. The practical implications of this study are as follows.
The drug type that victims use impacts survival rates. Mainly, fentanyl and heroin
decrease the survival odds, as compared to other drugs, such as marijuana. This result
implicates that policies are needed to eliminate highly addicted drugs from the community.
Additionally, safe sites providing a medically supervised environment in which drug addicts
can use the drug of their choice, particularly heroin, can be considered. The response time is
the most crucial variable, increasing survival rates; thus, safe sites can immediately respond
to the OD incidents. Moreover, these safe sites can provide mental health services and help
addicts reconnect with society. However, there is a risk that safe sites encourage opioid usage
and increase crime rates in their surroundings.
OD’ed victims’age determines his/her likelihood to survive. People younger than 40 have
higher survival rates than people older than 40. The primary goal should be to prevent OD,
particularly among older adults. Since older adults tend to have more health problems (e.g.
chronic pain), they are more likely to be prescribed opioids, causing them to get addicted.
Artificial
intelligence for
emergency

Governments can develop and manage programs designed for older adults to prevent them
from getting addicted to prescription/opioid drugs.
The time and day of the OD incident impact the survival rates. People who OD’ed earlier in
the day and during the weekend are more likely to survive. Government agencies managing
the emergency response team should consider the OD rates when allocating resources. For
example, the survival rates during earlier times in a day are much higher. This may indicate
that the resources are being scarce later in the day, delaying the response time.
Counties and regions that are more populated and considered urban have better survival
rates. This may be related to the revenues and resources of the counties. For example, urban
counties have better job opportunities, more business, lower unemployment rates and higher
median household incomes. Thus, these counties may have more tax revenue that can be
utilized to increase their resources (e.g. purchasing equipment and vehicles or hiring
emergency medical technicians and police officers), thus having a better response capability
to OD incidents.
There are some challenges related to the usage of such AI in practice. First, AI typically
requires large amounts of accurate data as input. These data can be available in some regions
of the world or for some communities, but not in other regions or for other communities due to
legal or technical reasons. For example, most of developed countries have documented data
detailed enough and sufficient in size to make AI predictions accurate, which may not be the
case for many developing countries. In some countries, even if data are available, some of
them may be inaccurate. Therefore, we recommend that the implementation of such
application must be accompanied by the establishment of policies and regulations for
appropriate data collection and management.
Second, we suggest that, given the unique environment of emergency management, the
use of AI needs to be more deliberate and careful so that proper attention can be paid to
aligning its usage to human values, such as optimizing the human well-being, and not as a
tool to reduce cost.
6.4 Limitations and future research
This study is not without limitations. One of them is the proposed study is implemented using
one source of data. Future research can apply our proposed AI-based solution to datasets
from different areas, states and countries. This will make the results more generalizable.
Another limitation resides in the fact this study focuses only on the first stage of the EMF –
the mitigation phase. Future studies can be conducted to improve the four phases of the EMF.
Additionally, future studies can combine this proposed methodology with optimization
techniques and simulation tools to develop resource allocation models.
7. Conclusion
This study examines how digital transformation –particularly open access digital data –AI
and ML can be utilized in emergency management. This study focuses on the opioid epidemic,
which impacts hundreds of thousands of people annually and costs trillions of dollars to the
worldwide economy. Using the design science research paradigm, this research integrates AI
with a well-established framework—EMF—to improve it by developing an explainable AI
solution that identifies opioid OD trends and determines the significant factors improving
survival rates. The proposed AI-based solution contains three stages. The first stage creates a
dataset from various open data sources, while the second phase trains three AI algorithms:
RF, ANN and SVM. The third phase utilizes SHAP to make the models more transparent and
generate insights into the significant factors affecting OD survival rates.
Governments and their policymakers can develop new strategies and allocate resources
by using this study’s results to fight and control this epidemic, which requires urgent
IMDS

attention. To this end, the findings discussed above provide a demonstration of how AI can
support informed decision-making in emergency management. Specifically, the explainable
AI-based solution shows an example of how AI can revolutionize the first stage of the EMF –
the mitigation phase. In providing government and emergency response agencies with
insights into the survival rates and resource allocation, the explainable AI-based solution also
helps in the second stage of the EMF –the preparation phase.
Notes
1. https://data.pa.gov/Opioid-Related/OD-Information-Network-Data-CY-January-2018-/hbkk-dwy3
2. https://data.pa.gov/Opioid-Related/Drug-and-Alcohol-Treatment-Facilities-May-2018-Cou/eswt-
bam9/data
3. https://data.pa.gov/Jobs-that-Pay/Pennsylvania-Registered-Corporations-Year-1768-Cur/g6yf-
7sah/data
4. https://data.pa.gov/Jobs-that-Pay/Local-Area-Unemployment-Statistics-LAUS-CY-2016-Cu/
rqq6-7e5m
5. https://www.census.gov/quickfacts/fact/table/PA/INC110218
6. https://data.pa.gov/Health/Enterprise-Data-Dissemination-Informatics-Exchange/aps5-fttf
7. https://data.pa.gov/Infrastructure-and-Development/Public-Transit-Services-and-Reduced-fare-
Programs-/wgkx-3uws
References
Aas, K., Jullum, M. and Løland, A. (2020), “Explaining individual predictions when features are
dependent: more accurate approximations to Shapley values”, available at: https://github.com/
NorskRegnesentral/shapr (accessed 4 June 2021).
Abbasi, A., Albrecht, C., Vance, A. and Hansen, J. (2012), “Metafraud: a meta-learning framework for
detecting financial fraud”,MIS Quarterly: Management Information Systems, University of
Minnesota, Vol. 36 No. 4, pp. 1293-1327.
Ancona, M., Cengiz€oztireli, C.C. and Gross, M. (2019), “Explaining deep neural networks with a
polynomial time algorithm for Shapley values approximation”,Proceedings of the 36th
International Conference on Machine Learning.
Badger, J., LaRose, E., Mayer, J., Bashiri, F., Page, D. and Peissig, P. (2019), “Machine learning forphenotyping
opioid overdose events”,Journal of Biomedical Informatics, Academic Press, Vol. 94, 103185.
Bellaire, S., van Herwijnen, A., Mitterer, C. and Schweizer, J. (2017), “On forecasting wet-snow
avalanche activity using simulated snow cover data”,Cold Regions Science and Technology,
Elsevier B.V., Vol. 144, pp. 28-38.
Bohr, A. and Memarzadeh, K. (2020), “The rise of artificial intelligence in healthcare applications”,in
Artificial Intelligence in Healthcare, Elsevier, pp. 25-60, available at: https://www.ncbi.nlm.nih.
gov/pmc/articles/PMC7325854/.
Boslett, A.J., Denham, A. and Hill, E.L. (2020), “Using contributing causes of death improves prediction
of opioid involvement in unclassified drug overdoses in US death records”,Addiction, Blackwell
Publishing, Vol. 115 No. 7, pp. 1308-1317.
Breiman, L. (2001), “Random forests”,Machine Learning, Vol. 45 No. 1, pp. 5-32.
Breiman, L. (2002), “Manual on setting up, using, and understanding random forests V3.1”, available
at: https://www.stat.berkeley.edu/∼breiman/Using_random_forests_V3.1.pdf (accessed 4
June 2021).
Canon, M.J., Satuito, A. and Sy, C. (2019), “Determining disaster risk management priorities through a
neural network-based text classifier”,Proceedings - 2018 International Symposium on
Artificial
intelligence for
emergency

Computer, Consumer and Control, IS3C 2018, Institute of Electrical and Electronics Engineers,
Taichung, pp. 237-241.
Carver, L. and Turoff, M. (2007), “Human-computer interaction: the human and computer as a team in
emergency management information systems”,Communications of the ACM, ACM PUB27 New
York, NY, Vol. 50 No. 3, pp. 33-38.
Cavdur, F. and Sebatli, A. (2019), “A decision support tool for allocating temporary-disaster-response
facilities”,Decision Support Systems, Elsevier B.V., Vol. 127, 113145.
CDC (2020), “Data overview: the drug overdose epidemic: behind the numbers”, available at: https://
www.cdc.gov/drugoverdose/data/index.html (accessed 14 December 2020).
Centers for Medicare and Medicaid Services (2018), “Ongoing emergencies”, available at: https://www.
cms.gov/About-CMS/Agency-Information/Emergency/EPRO/Current-Emergencies/Ongoing-
emergencies (accessed 4 June 2021).
Chaudhuri, N. and Bose, I. (2020), “Exploring the role of deep neural networks for post-disaster
decision support”,Decision Support Systems, Elsevier B.V., Vol. 130, 113234.
Chawla, N.V., Bowyer, K.W., Hall, L.O. and Kegelmeyer, W.P. (2011), “SMOTE: synthetic minority
over-sampling technique”,Journal of Artificial Intelligence Research, Vol. 16, pp. 321-357.
Cortes, C. and Vapnik, V. (1995), “Support-vector networks”,Machine Learning, Springer Science and
Business Media LLC, Vol. 20 No. 3, pp. 273-297.
Cutler, A., Cutler, D.R. and Stevens, J.R. (2012), “Random forests”,inEnsemble Machine Learning,
Springer US, Boston, MA, pp. 157-175.
De Maio, C., Fenza, G., Gaeta, M., Loia, V. and Orciuoli, F. (2011), “A knowledge-based framework for
emergency DSS”,Knowledge-Based Systems, Elsevier, Vol. 24 No. 8, pp. 1372-1379.
Delmolino, D. and Whitehouse, M. (2018), Responsible AI: A Framework for Building Trust in Your AI
Solutions.
Deryugina, T. (2017), “The fiscal cost of hurricanes: disaster aid versus social insurance”,American
Economic Journal: Economic Policy, American Economic Association, Vol. 9 No. 3, pp. 168-198.
Dong, X., Rashidian, S., Wang, Y., Hajagos, J., Zhao, X., Rosenthal, R.N., Kong, J., Saltz, M., Saltz, J. and
Wang, F. (2019), “Machine learning based opioid overdose prediction using electronic health
records”,AMIA Annual Symposium Proceedings, pp. 389-398, available at: https://www.ncbi.
nlm.nih.gov/pmc/articles/PMC7153049/.
Dou, M., Chen, J., Chen, D., Chen, X., Deng, Z., Zhang, X., Xu, K. and Wang, J. (2014), “Modeling and
simulation for natural disaster contingency planning driven by high-resolution remote sensing
images”,Future Generation Computer Systems, Elsevier, Vol. 37, pp. 367-377.
Dutta, I., Dutta, S. and Raahemi, B. (2017), “Detecting financial restatements using data mining
techniques”,Expert Systems with Applications, Elsevier, Vol. 90, pp. 374-393.
Erdoǧan, G., Erkut, E., Ingolfsson, A. and Laporte, G. (2010), “Scheduling ambulance crews for
maximum coverage”,Journal of the Operational Research Society, Palgrave Macmillan, Vol. 61
No. 4, pp. 543-550.
Fern
andez, A., Garc

ıa, S., Herrera, F. and Chawla, N.V. (2018), “SMOTE for learning from imbalanced
data: progress and challenges, marking the 15-year anniversary”,Journal of Artificial
Intelligence Research, Vol. 61, doi: 10.1613/JAIR.1.11192.
Fertier, A., Barthe-Delano€
e, A.M., Montarnal, A., Truptil, S. and B
enaben, F. (2020), “A new emergency
decision support system: the automatic interpretation and contextualisation of events to model
a crisis situation in real-time”,Decision Support Systems, Elsevier B.V., Vol. 133, 113260.
Florence, C., Luo, F. and Rice, K. (2021), “The economic burden of opioid use disorder and fatal opioid
overdose in the United States, 2017”,Drug and Alcohol Dependence, Vol. 218, 108350.
Gama, M., Santos, B.F. and Scaparra, M.P. (2016), “A multi-period shelter location-allocation model
with evacuation orders for flood disasters”,EURO Journal on Computational Optimization,
Springer Verlag, Vol. 4 Nos 3-4, pp. 299-323.
IMDS

Gul, M., Fuat Guneri, A. and Gunal, M.M. (2020), “Emergency department network under disaster
conditions: the case of possible major Istanbul earthquake”,Journal of the Operational Research
Society, Taylor and Francis, Vol. 71 No. 5, pp. 733-747.
Hayes, P. and Kelly, S. (2018), “Distributed morality, privacy, and social media in natural disaster
response”,Technology in Society, Vol. 54, Elsevier, pp. 155-167.
Haykin, S. (2009), Neural Networks and Learning Machines, 3rd ed., Prentice-Hall, Ontario.
Hevner, A.R., March, S.T., Park, J. and Ram, S. (2004), “Design science in information systems
research”,MIS Quarterly: Management Information Systems, Management Information
Systems Research Center, Vol. 28 No. 1, pp. 75-105.
Jamali, M., Nejat, A., Ghosh, S., Jin, F. and Cao, G. (2019), “Social media data and post-disaster
recovery”,International Journal of Information Management, Elsevier, Vol. 44, pp. 25-37.
James, G., Witten, D., Hastie, T. and Tibshirani, R. (2013), An Introduction to Statistical Learning,
Vol. 103, Springer, New York, NY, doi: 10.1007/978-1-4614-7138-7.
Johnson, M., Albizri, A. and Simsek, S. (2020), “Artificial intelligence in healthcare operations to
enhance treatment outcomes: a framework to predict lung cancer prognosis”,Annals of
Operations Research, Springer, pp. 1-31, doi: 10.1007/s10479-020-03872-6.
Johnson, M., Albizri, A. and Harfouche, A. (2021), “Responsible artificial intelligence in healthcare:
predicting and preventing insurance claim denials for economic and social wellbeing”,
Information Systems Frontiers, Springer, pp. 1-17, doi: 10.1007/s10796-021-10137-5.
Kapletia, D., Phillips, W., Medcalf, N., Makatsoris, H., McMahon, C. and Rich, N. (2019), “Redistributed
manufacturing–challenges for operations management”,Production Planning and Control,
Vol. 30 No. 7, pp. 493-495.
Kim, J.-M., Woods, P.K., Park, Y.J., Kim, T. and Son, K. (2016), “Predicting hurricane wind damage by
claim payout based on Hurricane Ike in Texas”,Geomatics, Natural Hazards and Risk, Taylor
and Francis, Vol. 7 No. 5, pp. 1513-1525.
Ko, B. and Kwak, S. (2012), “Survey of computer vision–based natural disaster warning systems”,
Optical Engineering, SPIE-Intl Soc Optical Eng, Vol. 51 No. 7, 070901.
Kohavi, R. (1995), “A study of cross-validation and bootstrap for accuracy estimation and model
selection”,Appears in the International Joint Conference on Artificial Intelligence (IJCAI), pp. 1-7.
Ladds, M., Keating, A., Handmer, J. and Magee, L. (2017), “How much do disasters cost? A comparison
of disaster cost estimates in Australia”,International Journal of Disaster Risk Reduction, Vol. 21,
pp. 419-429.
Lee, H.R. and Lee, T. (2021), “Multi-agent reinforcement learning algorithm to solve a partially-
observable multi-agent problem in disaster response”,European Journal of Operational
Research, Elsevier B.V., Vol. 291 No. 1, pp. 296-308.
Lo-Ciganic, W.H., Huang, J.L., Zhang, H.H., Weiss, J.C.,Wu,Y.,Kwoh,C.K.,Donohue,J.M.,Cochran,G.,
Gordon, A.J., Malone, D.C., Kuza, C.C. and Gellad, W.F. (2019), “Evaluation of machine-learning
algorithms for predicting opioid overdose risk among Medicare beneficiaries with opioid
prescriptions”,JAMA Network Open, NLM (Medline), Vol. 2 No. 3, e190968.
Lo-Ciganic, W.-H., Donohue, J.M., Hulsey, E.G., Barnes, S., Li, Y., Kuza, C.C., Yang, Q., Buchanich, J.,
Huang, J.L., Mair, C., Wilson, D.L. and Gellad, W.F. (2021), “Integrating human services and
criminal justice data with claims data to predict risk of opioid overdose among Medicaid
beneficiaries: a machine-learning approach”, in Zaller, N.D. (Ed.), PLOS ONE, Public Library of
Science (PLoS), Vol. 16 No. 3, e0248360.
Luna-Reyes, L.F. and Gil-Garcia, J.R. (2014), “Digital government transformation and internet portals:
the co-evolution of technology, organizations, and institutions”,Government Information
Quarterly, Elsevier, Vol. 31 No. 4, pp. 545-555.
Lundberg, S. and Lee, S.-I. (2017), “A unified approach to interpreting model predictions”,31st
Conference on Neural Information Processing Systems, available at: https://arxiv.org/pdf/1705.
07874.pdf (accessed 4 June 2021).
Artificial
intelligence for
emergency

McLoughlin, D. (1985), “A framework for integrated emergency management”,Public Administration
Review, JSTOR, Vol. 45, p. 172.
MedlinePlus - U.S. National Library of Medicine (2020), “Opioid addiction”, available at: https://
medlineplus.gov/genetics/condition/opioid-addiction/ (accessed 4 June 2021).
Mendonça, D.J. and Wallace, W.A. (2007), “A cognitive model of improvisation in emergency
management”,IEEE Transactions on Systems, Man, and Cybernetics Part A: Systems and
Humans, Vol. 37 No. 4, pp. 547-561.
Neill, D.B. and Herlands, W. (2018), “Machine learning for drug overdose surveillance”,Journal of
Technology in Human Services, Routledge, Vol. 36 No. 1, pp. 8-14.
NIH (2020), “Opioids”,National Institute on Drug Abuse, available at: https://www.drugabuse.gov/
drug-topics/opioids (accessed 4 June 2021).
Nissen, N. (2021), “Opioid overdoses 29% higher in 2020 than before the pandemic: study”, available
at: https://abcnews.go.com/Health/opioid-overdoses-29-higher-2020-pandemic-study/story?
id575785104 (accessed April 2021).
Noble, W.S. (2006), “What is a support vector machine?”,Nature Biotechnology, Nature Publishing
Group, Vol. 24 No. 12, pp. 1565-1567.
Olson, D.L. and Delen, D. (2008), Advanced Data Mining Techniques, Springer Publishing Company,
Incorporated, Berlin, doi: 10.1007/978-3-540-76917-0.
Pekar, V., Binner, J., Najafi, H., Hale, C. and Schmidt, V. (2020), “Early detection of heterogeneous
disaster events using social media”,Journal of the Association for Information Science and
Technology, John Wiley and Sons, Vol. 71 No. 1, pp. 43-54.
Petak, W.J. (1985), “Emergency management: a challenge for public administration”,Public
Administration Review, JSTOR, Vol. 45, p. 7.
Prieto, J.T., Scott, K., McEwen, D., Podewils, L.J., Al-Tayyib, A., Robinson, J., Edwards, D., Foldy, S.,
Shlay, J.C. and Davidson, A.J. (2020), “The detection of opioid misuse and heroin use from
paramedic response documentation: machine learning for improved surveillance”,Journal of
Medical Internet Research, JMIR Publications, Vol. 22 No. 1, e15645.
Ramachandran, P., Zoph, B. and Le, Q.V. (2017), “Searching for activation functions”,6th International
Conference on Learning Representations, ICLR 2018 - Workshop Track Proceedings,
International Conference on Learning Representations, ICLR, Vancouver.
Ramchurn, S.D., Huynh, T.D., Wu, F., Ikuno, Y., Flann, J., Moreau, L., Fischer, J.E., Jiang, W., Rodden,
T., Simpson, E., Reece, S., Roberts, S. and Jennings, N.R. (2016), “A disaster response system
based on human-agent collectives”,Journal of Artificial Intelligence Research, Vol. 57,
pp. 661-708.
Reddick, C. (2011), “Information technology and emergency management: preparedness and planning
in US states”,Disasters, John Wiley & Sons, Vol. 35 No. 1, pp. 45-61.
Roscoe, S., Cousins, P.D. and Handfield, R. (2019), “The microfoundations of an operational capability
in digital manufacturing”,Journal of Operations Management, John Wiley & Sons, Vol. 65 No. 8,
pp. 774-793.
Simsek, S., Albizri, A., Johnson, M., Custis, T. and Weikert, S. (2021), “Predictive data analytics for
contract renewals: a decision support tool for managerial decision-making”,Journal of
Enterprise Information Management, Vol. 34 No. 2, pp. 718-732, doi: 10.1108/JEIM-12-2019-0375.
Sun, W., Bocchini, P. and Davison, B.D. (2020), “Applications of artificial intelligence for disaster
management”,Natural Hazards, Springer, 1 September.
Tang, P. and Shen, G.Q. (2015), “Decision-making model to generate novel emergency response plans
for improving coordination during large-scale emergencies”,Knowledge-Based Systems,
Elsevier, Vol. 90, pp. 111-128.
The Office of Governor Tom Wolf (2016), “So what is open data?”, available at: https://www.governor.
pa.gov/newsroom/what-is-open-data-and-common-questions/ (accessed 4 June 2021).
IMDS

The United States Open Data (2018), “Open government”, available at: https://www.data.gov/open-
gov/ (accessed 4 June 2021).
TheGovLab (2020), “Open data demand: toward an open data demand assessment and segmentation
methodology”, available at: https://www.thegovlab.org/open-data-demand.html (accessed 4
June 2021).
Ti, L., Ho, A. and Knight, R. (2021), “Towards equitable AI interventions for people who use drugs:
key areas that require ethical investment”,Journal of Addiction Medicine, Ovid Technologies
(Wolters Kluwer Health), Vol. 15 No. 2, pp. 96-98.
Tinguaro Rodr
ıguez, J., Vitoriano, B. and Montero, J. (2010), “A natural-disaster management DSS for
humanitarian non-governmental organisations”,Knowledge-Based Systems, Elsevier, Vol. 23
No. 1, pp. 17-22.
Vial, G. (2019), “Understanding digital transformation: a review and a research agenda”,Journal of
Strategic Information Systems, Elsevier B.V., Vol. 28 No. 2, pp. 118-144.
Wamba, S.F. and Queiroz, M.M. (2020), “Blockchain in the operations and supply chain management:
benefits, challenges and future research opportunities”,International Journal of Information
Management, Vol. 52. 102064, ISSN 0268-4012, doi: 10.1016/j.ijinfomgt.2019.102064.
Ward, P.J., Rock, P.J., Slavova, S., Young, A.M., Bunn, T.L. and Kavuluru, R. (2019), “Enhancing
timeliness of drug overdose mortality surveillance: a machine learning approach”,inPławiak,
P. (Ed.), PLOS ONE, Public Library of Science, Vol. 14 No. 10, e0223318.
Waugh, W. and Streib, G. (2006), “Collaboration and leadership for effective emergency management”,
Public Administration Review, Vol. 66, doi: 10.1111/j.1540-6210.2006.00673.x.
WHO (2020), “Opioid overdose”,Blackwell Publishing, available at: https://www.who.int/news-room/
fact-sheets/detail/opioid-overdose (accessed 14 December 2020).
Yao, H., Rashidian, S., Dong, X., Duanmu, H., Rosenthal, R.N. and Wang, F. (2020), “Detection of
suicidality among opioid users on Reddit: machine learning-based approach”,Journal of
Medical Internet Research, JMIR Publications, Vol. 22 No. 11, e15293.
Zhang, J.H., Li, J. and Liu, Z.P. (2012), “Multiple-resource and multiple-depot emergency response
problem considering secondary disasters”,Expert Systems with Applications, Pergamon, Vol. 39
No. 12, pp. 11066-11071.
Zhang, C., Liu, X., Jiang, Y., Fan, B. and Song, X. (2016), “A two-stage resource allocation model for
lifeline systems quick response with vulnerability analysis”,European Journal of Operational
Research, Elsevier B.V., Vol. 250 No. 3, pp. 855-864.
Zheng, S. and Liu, H. (2019), “Improved multi-agent deep deterministic policy gradient for path
planning-based crowd simulation”,IEEE Access, Institute of Electrical and Electronics
Engineers, Vol. 7, pp. 147755-147770.
Zhu, L., Chen, P., Dong, D. and Wang, Z. (2021), “Can artificial intelligence enable the government to
respond more effectively to major public health emergencies? Taking the prevention and
control of Covid-19 in China as an example”,Socio-Economic Planning Sciences, 101029, ISSN
0038-0121.
Zubiaga, A., Aker, A., Bontcheva, K., Liakata, M. and Procter, R. (2018), “Detection and resolution of
rumours in social media: a survey”,ACM Computing Surveys, Association for Computing
Machinery, 1 February.
Corresponding author
Abdullah Albizri can be contacted at: albizria@montclair.edu
For instructions on how to order reprints of this article, please visit our website:
www.emeraldgrouppublishing.com/licensing/reprints.htm
Or contact us for further details: permissions@emeraldinsight.com
Artificial
intelligence for
emergency