Design of a modular DSS for public health decision-making in the context of a COVID-19 pandemic landscape.

The awareness of the occurrence of a new disease involves much uncertainty and the search for answers and also appropriate questions. In this paper we focus on the perspective of public health decision-makers. Typically, they would have a standard set of questions and supporting metrics that have been found in previous disease outbreaks to be useful in assessing the effectiveness of various 'solution' methods on the trajectory of the disease. There may be other relevant questions with which such public health domain experts may not be familiar and/or for which they are familiar but are not aware of methods for addressing such questions when there is limited data. Decision Support Systems (DSS) can be used to facilitate the exploration of established questions and some other relevant questions. Given an initial set of questions, the DSS designer should consider which sets of data analytic methods have the capabilities to adequately address. Some of these data analytic methods may also have the capability of addressing questions that could be of interest to the public health decision makers including researchers. In this paper we present a conceptual design for a relevant easy-to-construct DSS and an example of a multi-method DSS that is based on this conceptual design. Using publicly available data on the CoViD-19 pandemic, we illustrate benefits of the multi-method DSS in action.

Introduction

A process for dealing with a pandemic, or with the spread of any disease, could be seen as being comprised of two distinctive, yet interrelated parts that are combined into a system characterized, as a result, by a high cohesion and loose coupling of its components. The first component would be a general decision-making model (e.g. Samoilenko and Osei-Bryson, 2013, Babaei and Bamdad, 2020, Shi et al., 2017) that is customizable to the context of a given pandemic via the information provided by the second model. Unlike the first component, the second model is a specific to a given disease decision support system (DSS) that would be primarily utilized and relied upon by the relevant public health domain decision makers.

The field of information systems (IS) is uniquely and advantageously positioned to help fighting spread of a disease via designing and developing a DSS (Collins, Ketter & Gini, 2010) that could support public health decision makers in a pandemic. The purpose of this paper is to outline a design of such a DSS. We accomplish this goal via three objectives. First, we offer a conceptual design of a DSS based on the set of explicitly stated assumptions and premises. Second, we offer a blueprint of the system by designing an appropriate multi-method methodology that relies on the tools of data analysis and data mining that are commonly used in the field of IS. Finally, we test the resultant decision support system in the context of the current pandemic using publicly available data.

Our effort is suggestive, rather than definitive, in its purpose, for we are not claiming that our approach is the only way that the DSS could be designed and constructed. Instead, we invite our readers to contribute their ideas and to use our effort as a possible foundation in creating and testing better designs.

We present our work in the following sequence: Section 2 describes conceptual design of the DSS. Section 3 outlines the methodology that implements the design, and 4, 5 are focused on the testing of the system using publicly available data. Conclusion and Discussion’ sections close the paper.

Conceptual design of the DSS

Any conceptual design of an artifact relies on a particular model of the environment within which the artifact is intended to operate. We rely on a set of assumptions (A#) that serve the purpose of the describing the environment within which a disease spreads. We invite our reader to examine each of the assumptions, along with the corresponding justifications, presented below in Table 1 .

Table 1

Conceptual Design of a DSS: Underlying Assumptions and Justifications.

Assumption		Justification
A1	For each disease, there is an identifiable set of demographic risk factors that provides a useful characterization of the disease.	For each known disease, public health practitioners attempt to identify variables that have high correlations with its occurrence and outcomes. For each disease, the identified variables are its ‘risk factors’. Once these ‘risk factors have been identified public health agencies aim to collect data that can be used to estimate that associated population-level statistics.
A2	The population of a geographic area (e.g., nation, state, city, county, etc.) could be described by a set of demographic factors that includes the risk factors for known diseases.	In practice this typically holds at the national and state levels, and in some cases at the county and city level.
A3	The population of a geographic area could be characterized by a disease-related subset of its demographic factors i.e. the risk factors for a given disease.	In practice for known diseases this typically holds at the national and state levels, and in some cases at the county and city level.
A4	Populated geographic areas could be grouped in terms of the risk factors for a given disease.	If A3 holds then A4 should also be possible.
A5	A geographic area could be assessed in terms of the impact of the risk factors for a given disease on the associated infection rate.	The reasonableness of this assumption follows from the meaning of the concept of a ‘risk factor’ for a disease.
A6	The spread of an infectious disease follows a path towards geographic areas with higher demographic risk factors.	The reasonableness of this assumption follows from the meaning of the concept of a ‘risk factor’ for a disease.
A7	A geographic area could be assessed in terms of its relative effectiveness and efficiency of containing the spread of a disease.	The presence of any organized system of healthcare is associated with collecting of the relevant patient data.
A8	A spread of an infectious disease follows a path towards geographic areas with lower levels of effectiveness and efficiency of containing the spread of a disease.	The reasonableness of this assumption would follow from the meaning of the concepts of ‘effectiveness’ and ‘efficiency’ of disease containment.
A9	A geographic area could be assessed in terms of the changes in its level of efficiency of containing a disease over time.	If A7 holds then A9 would also hold.
A10	The efficiency and effectiveness of a geographic area in containing a disease could be improved via area-specific decisions (e.g. allocation of resources).	If A10 does not hold then it would not make sense to have a DSS that supports the making of appropriate area-specific decisions.

Conceptually, the design of the DSS could be perceived as consisting of two modules. The first module allows for projecting the direction/flow of contagion across geographic areas based on the demographic factors. The second module allows for modeling the spread of the disease based on the effectiveness and efficiency of the geographic areas in fighting the outbreak. The operation of the modules is supported by the following two propositions (P#):

P1: A preferred path of the spread of a disease is toward geographic area(s) characterized by high-risk demographic factors specific to the disease.

P2: A path of contagion is directed towards the geographic area(s) that is (are) least efficient and effective in containing the spread of the disease.

In order to defend the propositions, we use the rules of hypothetico-deductive logic so our reader can examine veracity of the statements.

In regard to P1, the supporting argument is as follows:

: A spread of a disease X is associated with a set of demographic risk factors

: Area A has a greater proportion of the population with a set of X-specific risk factors that are at a higher level than that of Area B

: Area A will have a greater level of contagion of X than Area B.

The supporting argument for P2 is as follows:

: A spread of a disease X is associated with a level of efficiency and effectiveness of containment of the outbreak

: Area A has a lower level of efficiency and effectiveness of containing X than Area B

: Area A will have a greater level of contagion of X than Area B.

The above-stated set of assumptions allows for outlining a set of steps that the proposed DSS should allow a decision maker for performing. This set of steps is grouped into two modules, where each module is would be applied at a different state/stage of a pandemic.

Module 1: Identify a path of contagion based on demographic risk factors

State of the pandemic: Initial stage, pre-pandemic, patient zero to limited number of cases.

Step 1: Identify a set of demographic risk factors for the given disease. The outcome of this step is a disease-specific sub-set of the variables that are used to describe the population of interest.

Step 2: Test the assumption of homogeneity of geographic areas of interest by using the available demographic data. The outcome of this step is a set of groups of geographic areas that may differ in terms of the general demographic factors.

Step 3: Identify the demographic variables that differentiate the set of groups of geographic areas the most. The outcome of this step is a sub-set of the variables that are, possibly, also risk factors for a given disease.

Step 4: Discover naturally occurring associations between demographic risk factors and the level of contagion. The outcome of Step 4 is a confirmation of the relationship between the demographic risk factors and the contraction of the disease.

Step 5: Test for the presence of the impact of the demographic risk factors on the spread of the disease. The result would allow for identifying causal impact of the risk factors on the contraction of the disease.

The output of using Module 1 is an n-tiered “projection of a path” system comprised of, if n = 3, Low Risk, Mid Risk, and High Risk areas and constructed using the actionable information obtained in Step 5.

Module 2: Identify a path of contagion based on efficiency and effectiveness of containing the spread

State of the pandemic: Developing stage, pandemic is growing, number of cases is rising.

Step 6: Identify groups of geographic areas based on the effectiveness of containing the spread of the disease. This would allow for ranking the geographic areas in terms of the relative success in dealing with the given disease.

Step 7: Compare the groups identified in Step 6 with the groups identified in Step 2.

Step 8: Identify the factors differentiating the areas discerned as a result of Step 6.

Step 9: Assess a relative efficiency of containing the spread of the disease of a geographic area vis-à-vis other areas. The outcome of this step is a ranked order of areas reflecting their relative standing of each area vis-à-vis the better and worse performing counterparts.

Step 10: Assess the changes, over time, in the relative efficiency of each geographic area in regard to dealing with the contagion. The outcome of this step is an assessment, of each area, in regard to its improvement/deterioration of its performance in fighting the disease that took place over the given time period(s).

Step 11: Identify the drivers of the change in performance of a geographic area in fighting the disease. The outcome of this step is the discernment of the culprit of the improving or deteriorating performance.

The output of using Module 2 is an identification of High, Mid, and Low risk areas based on the efficiency and effectiveness in containing the spread of the diseases (and minimizing the mortality rate) with a correspondent suggestion for the improvements.

At this point we are ready to present a blueprint of the DSS supported by the tools of the data analysis and data mining commonly available to the researchers and practitioners in the field of IS.

Design of the DSS based on an integrated multi-method workbench

The process of translation of a conceptual model into a blueprint of the design of DSS could follow two distinct paths. First option is to rely on a custom solution that is subject- and purpose-specific. This is akin to creating a program/application by relying on custom, one-off code. Such an option, while not without its merits, is characterized by the difficulties in maintaining and adapting the design when changes are called for. The second option is to opt for a “building block” approach, where the translation of the concept into the blueprint is achieved via existing and tried-and-true components. This is similar to creating a program by utilizing existing libraries of algorithms, data structures, and classes. As a result, the design is modular, highly transparent, and adaptable.

We follow the second approach in translating a conceptual model of DSS into the design blueprint, where established data analysis methods serve as the elements comprising the end product. We believe that the selected “building blocks” approach would allow our reader to not only evaluate the appropriateness and fit for the purpose of each component, but also to appraise the soundness of the overall design of the DSS. While we do not claim that the selected methods (i.e. Cluster Analysis, Decision Tree induction, Association Rules Mining, Data Envelopment Analysis, Multiple Regression) are the only and the best options for designing a DSS, we do suggest that the chosen methods are appropriate and well-suited for the purpose. We invite our reader to consider possible substitutes to the selected methods that may contribute to a more flexible and robust design of a DSS.

We offer a brief overview of the insights offered by each method, along with some of the limitations, in Table 2 below. For further details on these data analytic methods the reader could consult various studies (e.g. Samoilenko and Osei-Bryson, 2017, Osei-Bryson and Ngwenyama, 2014).

Table 2

Structural components of the DSS: offered insights and limitations.

Method	Offered Insight	Limitation
CA: Cluster Analysis	Allows for testing an assumption of homogeneity of the sample and identifying presence of sub-groups in the sample.	In the presence of multiple sub-groups does not offer any insights into the sources of heterogeneity.
DTI: Decision Tree Induction	Given the target variable, allows for identifying attributes responsible for differentiating sub-groups of the sample.	Target variable must be provided “from outside.” Does not consider impact of differentiating variables on an “Input → Output” model of sub-groups.
ARM: Association Rules Mining	Allows for identifying a set of “If ⇒ Then” rules present in the data set.	Does not provide any insights regarding an “Input → Output” process.
DEA: Data Envelopment Analysis	Allows for calculating the relative efficiency scores of decision-making units (DMUs), as well as changes in the scores over time via using the Malmquist Index (MI) scores.	A “black box” model of the “Input → Output” conversion process. Does not offer insights into the sources of inefficiencies.
MR: Multiple Regression	Allows for determining the significance of the impact of independent variables on a dependent variable and identifying the presence of complementarities.	Does not provide any insights regarding an “Input → Output” process and does not allow for considering multiple outputs of the process.

At this point we are ready to map the selected methods to the steps of each module. We offer our reader to examine each step, along with the intended results of the application of each method, by referring to Table 3 below.

Table 3

Modular Design of the DSS: Methodological Steps.

Module	Step	Method	Expected Outcome/Result
1	Step 2: Test the assumption of homogeneity of geographic areas of interest by using the available demographic data.	CA	A group of n-clusters of geographic areas that differ in terms of the demographic risk factors.
	Step 3: Identify the demographic variables that differentiate the geographic areas the most	DTI	A set of demographic factors that are responsible for the differences between the geographic areas.
	Step 4: Discover naturally occurring associations between demographic risk factors and the level of contagion.	ARM	A set of “if->then” naturally occurring associations that characterize the sample.
	Step 5: Test for the presence of the impact of the demographic risk factors and the spread of the disease.	MR	Determination of the significance/presence of the impact between “if” and “then” parts of associations discovered in Step 4.
2	Step 6: Identify groups of geographic areas based on the effectiveness of containing the spread of the disease.	CA	A group of n-clusters of geographic areas that differ in terms of the contagion-specific factors.
	Step 8: Identify the factors differentiating the areas discerned as a result of Step 6.	DTI	A set of contagion-specific factors that are responsible for the differences between the geographic areas identified in Step 8.
	Step 9: Assess a relative efficiency of containing the spread of the disease of a geographic area vis-à-vis other areas.	DEA	A set of scores of relative efficiency for each area, as well as for each cluster that was identified in Step 6.
	Step 10: Assess the changes, over time, in the relative efficiency of each geographic area in regard to dealing with the contagion.	DEA MI	Determination of the improvement, or deterioration of performance of each area in fighting the outbreak via Malmquist Index (MI) scores.
	Step 11: Identify the drivers of the change in performance of a geographic area in fighting the disease.	DEA MI	Determination of the reasons for the improvement/deterioration in performance of each geographic area via the relationship between the Efficiency Change component (EC) and Technology Change component (TC) of the MI scores.

At this point we are ready to outline the sequence of the methodological steps, for each module, as well as corresponding data flows in a pictorial format (see Fig. 1 ). Despite that our system is intended to be comprised into a coherent whole, we present our DSS as a collection of two loosely coupled modules, so our readers could evaluate each module independently. Also, we invite our readers to consider a suitability of using other methods of data analysis than those selected by us (e.g., to replace DEA with Free Disposal Hull (FDH), to substitute MR with multivariate adaptive regression splines (MARS), etc.).

Fig. 1

Design of the Proposed DSS.

The five techniques of data mining and data analysis that we use in the design of the proposed DSS have been widely utilized in IS research and practice in a stand-alone fashion. However, they are also very frequently used in combination to construct multi-method methodologies.

For example, DEA is widely employed for the purpose of evaluating productivity and performance (e.g. Khouja, 1995, Shao and Lin, 2001, Samoilenko and Green, 2008, Bollou and Ngwenyama, 2008, Yu and Lin, 2008, Avkiran and Rowlands, 2008, KAO and HUNG, 2008, Du et al., 2010, Lozano-Vivas and Pastor, 2010, Tsolas et al., 2020), but it has also been used to complement other data analytic techniques: cluster analysis (e.g. Shin and Sohn, 2004, Hirschberg and Lye, 2001, Lemos et al., 2005, Morais and Camanho, 2011), neural network induction (e.g. Samoilenko and Osei-Bryson, 2008, Çelebi and Bayraktar, 2008, Emrouznejad and Shale, 2009, Mostafa, 2009, Wu, 2009), decision tree induction (e.g. Samoilenko & Osei-Bryson, 2007; Samoilenko, 2008b; Wu, 2009), regression analysis (e.g. Cooper and Tone, 1997, Bollou and Ngwenyama, 2008, Parthasarathy and Anbazhagan, 2008, Samoilenko and Osei-Bryson, 2008), and other methods (Liu and Lu, 2010, Eilat et al., 2008, Ramanathan and Yunfeng, 2009).

Now, once a completed design of the system has been presented to our reader, we are ready to test the proposed DSS in action using relevant real world data.

Descriptions of the illustrative datasets

The context of the testing of the DSS is the United States, consequently, the required for the first module data were obtained from the United States Census Bureau (https://www.census.gov/acs/www/data/data-tables-and-tools/data-profiles/). The topic of interest is demographic data, which is available by selecting “Demographic Characteristics” option (https://data.census.gov/cedsci/table?d=ACS%205-Year%20Estimates%20Data%20Profiles&table=DP05&tid=ACSDP5Y2018.DP05). The latest available year for American Community Survey’ demographic and housing 5-year estimates is 2018, consequently, this year was selected. The data set was augmented by adding a variable “state’ Population Density” (https://state.1keydata.com/state-population-density.php), because population density is an important factor impacting the spread of a disease. We reduced the data set by selecting only those variables that are considered to be associated with Covid-19 risk factors (https://www.cdc.gov/coronavirus/2019-ncov/need-extra-precautions/index.html). Overall, we ended up with 50 variables (Table 4 ), available for 33 states (Table 5 ).

Table 4

Census Data - Variables Used in Module 1.

Variable/Code	Description
PopDensity	Population Density
S2601C_C01_010E	Total Population,55 to 64 years
S2601C_C01_011E	Total Population,65 to 74 years
S2601C_C01_012E	Total Population,75 to 84 years
S2601C_C01_013E	Total Population, 85 years +
S2601C_C01_017E	Total Population, 65 years +
S2601C_C01_018E	Total Population, 65 years +, Male
S2601C_C01_019E	Total Population, 65 years +, Female
S2601C_C01_020E	Total Population, Median age (years)
S2601C_C01_023E	Total Population, Black or African American
S2601C_C01_034E	Total population, 15 years +, Widowed
S2601C_C01_035E	Total population, 15 years +, Divorced
S2601C_C01_043E	Total population, 25 years +, Bachelor's degree or higher
S2601C_C01_047E	Total population With a disability
S2601C_C01_051E	Total Population 18 to 64 years With a disability
S2601C_C01_054E	Total population 65 years + With a disability
S2601C_C01_087E	Total population, 16 years +, Unemployed
S2601C_C01_088E	Total population 16 years +, Unemployed, Percent of civilian labor force
S2601C_C01_093E	Total population16 years +, Service occupations
S2601C_C01_106E	Total population, poverty rate, All people
S2601C_C01_107E	Total population, poverty rate, 18 years +
S2601C_C01_108E	Total population, poverty rate, 18 to 64 years
S2601C_C01_109E	Total population, poverty rate, 65 years +
S2601C_C02_009E	Total group quarters population, 45 to 54 years
S2601C_C02_010E	Total group quarters population, 55 to 64 years
S2601C_C02_011E	Total group quarters population, 65 to 74 years
S2601C_C02_012E	Total group quarters population, 75 to 84 years
S2601C_C02_013E	Total group quarters population, 85 years +
S2601C_C02_017E	Total group quarters population, 65 years +
S2601C_C02_018E	Total group quarters population, 65 years +, Male
S2601C_C02_019E	Total group quarters population, 65 years +, Female
S2601C_C02_023E	Total group quarters population, Black or African American
S2601C_C02_034E	Total group quarters population, 15 years +, Widowed
S2601C_C02_035E	Total group quarters population, 15 years +, Divorced
S2601C_C02_043E	Total group quarters population, 25 years +, Bachelor's degree or higher
S2601C_C02_047E	Total group quarters population, With a disability
S2601C_C02_051E	Total group quarters population, 18 to 64 years, With a disability
S2601C_C02_052E	Total group quarters population, 18 to 64 years, No disability
S2601C_C02_053E	Total group quarters population, Disability Status, 65 years +
S2601C_C02_054E	Total group quarters population, 65 years +, With a disability
S2601C_C02_087E	Total group quarters population,16 years +, Unemployed
S2601C_C02_088E	Total group quarters population,16 years +, Unemployed, % of the labor force
S2601C_C02_090E	Total group quarters population,16 years +, Not in labor force
S2601C_C02_093E	Total group quarters population, 16 years +, Service occupations
S2601C_C02_094E	Total group quarters population,16 years +, Sales and office occupations
S2601C_C02_105E	Total group quarters population, Individuals With Food Stamp/SNAP benefits
S2601C_C02_106E	Total group quarters population, Poverty Status is Determined, All people
S2601C_C02_107E	Total group quarters population, Poverty Status is Determined, 18 years +
S2601C_C02_108E	Total group quarters population, Poverty Status is Determined, 18 to 64 years
S2601C_C02_109E	Total group quarters population, Poverty Status is Determined, 65 years +

NB: Name of the State is also included as the ID variable.

Table 5

List of States Used in the Study.

Alabama, Arizona, Arkansas, California, Colorado, Connecticut, Florida, Georgia, Illinois, Indiana, Iowa, Kansas, Kentucky, Louisiana, Maryland, Massachusetts, Michigan, Minnesota, Mississippi, Missouri, New Jersey, New York, North Carolina, Ohio, Oklahoma, Oregon, Pennsylvania, South Carolina, Tennessee, Texas, Virginia, Washington, Wisconsin

The second module requires a disease-specific data, and Covid-19 data were obtained from COVID-19 Data Repository by the Center for Systems Science and Engineering (CSSE) at Johns Hopkins University (https://github.com/CSSEGISandData/COVID-19). For Steps 6–9 we selected the latest data available at the time of this stage of the study (07-15-2020), and for Steps 10 and 11 we used the data for April 12, May 13, June 14, July 15, and August 13 (data for August became available later in the study). Overall, this resulted in 4 time-periods of April-May, May-June, June-July, and July-August. The selected variables described in Table 6 .

Table 6

Pandemic Data – Variables Used in Module 2.

Variable	Description
Confirmed	Aggregated confirmed case count for the state.
Deaths	Aggregated Death case count for the state.
Active	Aggregated confirmed cases that have not been resolved.
Incidence_Rate	Confirmed cases per 100,000 persons.
People_Tested	Total number of people who have been tested.
Mortality_Rate	Number recorded deaths / Number confirmed cases.
Testing_Rate	Total number of people tested per 100,000 persons.

NB: Name of the State is also included as the ID variable.

Additionally, we used two variables to serve as a proxy for the level of medical resources available to each state: Total Active Patient Care Physicians, Rate per 100,000 (data were obtained from 2019 State Physician Workforce Data Report published by Association of American Medical Colleges, available on line at https://www.aamc.org/data-reports/workforce/report/state-physician-workforce-data-report) and Number of Hospitals (data were obtained from American Hospital Directory, available on line at https://www.ahd.com/state_statistics.html). Once the data sets were compiled, we tested the DSS in action- the results are described in the next section.

Testing the DSS-results of the data analysis

We present the results following the format that we used in describing the DSS- as a sequence of two modules, and the sequence of the steps within each module.

We used the input data set described in Table 4.

We used the RGui (R Core Team (2020). R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. URL https://www.R-project.org/.) data analysis to perform cluster analysis. Prior to performing CA, we standardized the imported dataset using scale function available in the package. Then we ran k-means CA with k = 3. The choice of k allowed us to partition the sample in terms of “High”, “Mid”, and “Low” risk areas- the determination of the label was based on the averaging of the rankings (in terms of the confirmed cases) of the members of each cluster (see Table 7.1b). The results are presented in Table 7.1a below.

Table 7.1b

Module 1 – Summary Descriptions of the Clusters.

Cluster	Cluster 1	Cluster 2	Cluster 3
Average rank	12.54	16.09	28.22
Assigned label	High Risk	Mid Risk	Low Risk
Cluster size	13	11	9

Table 7.1a

Module 1 – Cluster Membership.

Cluster	States
1	Arizona, California, Florida, Georgia, Illinois, Indiana, Michigan, New York, North Carolina, Ohio, Oregon, Pennsylvania, Texas
2	Colorado, Connecticut, Iowa, Kansas, Maryland, Massachusetts, Minnesota, New Jersey, Virginia, Washington, Wisconsin
3	Alabama, Arkansas, Kentucky, Louisiana, Mississippi, Missouri, Oklahoma, South Carolina, Tennessee

Once we got the memberships of each cluster, we could use it to calculate the average ranking of each cluster based on the nation-wide ranking of each state based on the Number of Confirmed Cases in descending sequence. The calculation demonstrates that average rank of Cluster 1 is higher than that of Cluster 2, and significantly higher than that of Cluster 3 (see Table 7.1b below). Interestingly, the 10 states comprising Cluster 1 (with the exception of Arizona, North Carolina and Oregon) are in the top 16 states in terms of the Number of Confirmed Cases.

By running decision tree analysis in RGui (R Core Team, 2017), we were able to identify two top-split variables that differentiate the clusters the most. After that we removed the two variables from the data set and re-ran the analysis with the purpose of identifying other variables that may play the role in differentiating the clusters. Results are presented in Table 7.1c and Fig. 2 below.

Table 7.1c

Module 1 – Top Variables that differentiate the Clusters.

DTI Iteration	Selected Splitting Variables
1	• Total population for whom poverty status is determined (PopPoverty) • Total population with a disability (PopDisability)
2	• Population Density (PopDensity) • Population 25 years and over with bachelor’s degree or higher (PopBS)

Fig. 2

Visual Representation of the Results of DTI.

In order to do the ARM analysis, we needed to augment the data set containing the demographic data with the data reflecting the number of confirmed cases. After that the data set was transformed so a record for every state could be represented as a transaction. We binned the numeric values for each variable into four quartiles- this allowed us to represent the range of values in terms of “Low”, MidLow”, MidHigh”, and “High” categories. After that we ran the analysis – the results are summarized in Table 7.1d below.

Table 7.1d

Module 1 – Association Rules.

Left Side (If)	Association (⇒)	Right Side (Then)
High Population Density	⇒	MidHigh Number of Confirmed Cases
High Population Density	⇒	High Number of Confirmed Cases
Low Population Density	⇒	LowMid Number of Confirmed Cases
Low Population Density	⇒	Low Number of Confirmed Cases
MidHigh Age 85+	⇒	High Number of Confirmed Cases
Low Poverty Level	⇒	LowMid Number of Confirmed Cases
MidLow Total Population with Disability	⇒	High Number of Confirmed Cases
Low Total Population with Disability	⇒	High Number of Confirmed Cases
High Total Population with Disability	⇒	Low Number of Confirmed Cases

The results of the previous step yielded a number of associations, but ARM does not allow for testing a significance of a causal relationship between the “If” and “Then” side of associations. So, in order to test the significance of the impact we use MR. Based on the results of ARM we created the following model:

(Age 85+, PopDensity, PopDisability, PopPoverty) → Number of Confirmed Cases.

We did not consider the interaction terms in the model.

The results of the analysis yielded Population Density, Population with Disability, and Population with Poverty to be statistically significant, while Population Age 85 + is not statistically significant. As a result, we removed that variable from the model and ran the following regression:

(PopDensity, PopDisability, PopPoverty) → Number of Confirmed Cases.

The results are summarized in the Table 7.1e below.

Table 7.1e

Module 1 – Result of MR Analysis.

Model Statistics	R2	Adjusted R2	F	Significance F
	0.390	0.327	6.190	0.002
Variable	Coefficients	Standard Error	t Stat	P-value
PopDensity	107.935	40.872	2.641	0.013
PopDisability	−20478.073	7514.452	−2.725	0.011
PopPoverty	13467.399	5916.581	2.276	0.030

Interestingly, the coefficient of Population with Disability is negative, which means that the states with a greater number of people with disability have a smaller number of confirmed cases. This may, possible, have to do with the limited mobility that some of the disabled people have, where the limited mobility implies limited exposure to others and, thus, constraining impact on the spread of the disease. At this point we have identified two variables- Population Density and Population Poverty- that have an impact on the spread of the disease (proxied by the Number of Confirmed Cases), and, consequently, we are in a good position to construct a projection of the path of a pandemic.

Pandemic path projection-prioritizing intervention measures

In projecting a path of a pandemic, knowing that Population Density and Population Poverty impact the Number of Confirmed Cases, we need to consider how actionable the information is. For example, while we can fairly quickly impact Population Density via social distancing, quarantines, and other restrictions, we cannot equally quickly impact Population Poverty. Consequently, let us consider only one identified variable, Population Density, as a predictor of the path of the spread of the disease. Based on the ranking of the states in regard to the population density, we could create a 3-tier projection system, where Tier 1 signified top 33% of the sample in terms of the population density, Tier 2 is comprised of the middle third, and Tier 3 is represented by the states in the bottom third of the sample. The results are summarized in Table 7.1f below.

Table 7.1f

Module 1 – Priority-based Groupings of the States.

Tier 1′ States- High Priority	Tier 2′ States- Mid Priority	Tier 3′ States- Low Priority
New Jersey, Massachusetts, Connecticut, Maryland, New York, Florida, Ohio, Pennsylvania, California, Illinois, Virginia	North Carolina, Indiana, Georgia, Michigan, South Carolina, Tennessee, Kentucky Washington, Texas, Wisconsin, Louisiana	Alabama, Missouri, Minnesota, Arizona, Mississippi, Arkansas, Oklahoma, Iowa, Colorado, Oregon, Kansas

The accuracy of the projection is easy to test, for the ranking of the states in terms of the number of cases is available- 11 of Tier 1′ states are in the top-13 states with respect to Number of Confirmed Cases (top-14 if counting Louisiana, but this state is not included in our data set).

In this step we perform k-means CA using the data set described in Table 6 (i.e. Pandemic).

The results are presented in Table 7.2a below.

Table 7.2a

Module 2 – Results of Cluster Analysis.

Cluster	States
1 (n = 13)	Colorado, Indiana, Kansas, Kentucky, Michigan, Minnesota, Missouri, Ohio, Oklahoma, Oregon, Pennsylvania, Washington, Wisconsin
2 (n = 12)	Alabama, Arizona, Arkansas, Georgia, Iowa, Maryland, Mississippi, North Carolina, South Carolina, Tennessee, Texas, Virginia
3 (n = 8)	California, Connecticut, Florida, Illinois, Louisiana, Massachusetts, New Jersey, New York

In this step we compare the results of the CA of Module 1 with the results of the previous step (i.e. CA of Module 2). The purpose of this comparison is to see how closely the contents of the clusters match- in a perfect world with the absence of inefficiencies, we would expect that in both cases the clusters would be comprised of the same states. Meaning, a cluster with a higher level of demographic risk factors would also be a cluster with a greater level of the spread of the disease (Table 7.2b ), and vice versa.

Table 7.2b

Module 2 – Summary Descriptions of the Clusters.

Cluster	Cluster 1	Cluster 2	Cluster 3
Average Rank	23.00	20.50	6.13
Assigned Label	Low Prevalence	Mid Prevalence	High Prevalence
Cluster Size	13	12	8

As previously, we used the ranking of the states in terms of the Number of Confirmed Cases to determine the averaged rank for each cluster. The determination of the label (i.e. High, Mid, Low) was based on the averaged rank.

Overall, we found out that 14 states mapped to the cluster perfectly- meaning, 42% of the sample exhibited an “expected behavior” (e.g. if they were in the High Risk category in Module 1’ cluster they were also in the High Prevalence category of Module 2’ cluster), 7 states, or 21%, did worse, while 12 states, or 36%, performed better than expected based on the pandemic data. The results are summarized in Table 7.2c .

Table 7.2c

Module 2 – Comparison of Cluster Memberships.

Confirmed Cases	High Level	Mid Level	Low Level
Based on demographics	Module 1: Cluster 1	Module 1: Cluster 2	Module 1: Cluster 3
Based on actual spread	Module 2: Cluster 3	Module 2: Cluster 2	Module 2: Cluster 1
Change in Avg. Ranking	12.54 → 6.12	16.09 → 20.5	28.22 → 23
States “As Expected” (n = 14, 42% of the sample)	o California, Florida, Illinois, New York	o Colorado, Kansas, Minnesota, Washington, Wisconsin	o Alabama, Arkansas, Mississippi, South Carolina, Tennessee
States that “Do Worse” (n = 7, 21% of the sample)	From Low to High:	From Mid to High:	From Low to Mid:
	o Louisiana	o Connecticut, Massachusetts, New Jersey	o Oklahoma, Missouri, Kentucky
States that “Do Better” (n = 12, 36% of the sample)	From High to Mid:	From Mid to Low:	From High to Low:
	o Ohio, Oregon, Pennsylvania, Indiana, Michigan	o Virginia, Maryland, Iowa	o Arizona, Texas, Georgia, North Carolina

It is worth noting that eight members of High Prevalence’ Cluster 3 are among the top 11 states (top-12 if counting Louisiana that is not a part of our data set) in terms of the Number of Confirmed Cases.

By performing classification DTI, we can identify the top-level splits responsible for the separation of the sample into the sub-groups. The variables identified were Incidence Rate and Deaths. We present the results in the Table 7.2d below.

Table 7.2d

Module 2 – Top Variables that differentiate the Module 2 Clusters.

Cluster	Condition
Cluster 1 (Low Prevalence - Lower level of spread)	Incidence Rate < 819.9
Cluster 2 (Mid Prevalence -Middle level of spread)	Incidence Rate ≥ 819.9 AND Deaths < 3403.5
Cluster 3 (High Prevalence -Higher level of spread)	Incidence Rate ≥ 819.9 AND Deaths ≥ 3403.5

The results suggest, specifically in regard to the difference between Cluster 2 and Cluster 3, that there are context-specific inefficiencies that we may want to look at. We’ll do in the next step.

For this step we created and ran three DEA models.

The first model allows us to assess a relative efficiency of each state, as well as the average for each cluster, in “converting” the demographic risk factors into cases. The model that we used is “(Population Density, Population with Disability, Population Poverty) → Number of Cases”. We ran input-oriented DEA (initial conditions are controlled and Number of Cases is to be manipulated) under assumption of variable return to scale. Due to the nature of output-oriented DEA, which aims at maximization of the output, we inverted the output variable (e.g., Number of Cases) by subtracting, for each state, the actual number of cases from 400000. This allows us to run a model that “rewards” decision making units with the smaller, rather than larger, number of cases.

The purpose of second and third DEA models was to test the relative efficiency of the states in terms of utilization of the available medical resources, which were represented by two proxy variables: Total Active Patient Care Physicians, Rate per 100,000 and Number of Hospitals. The second DEA model has as its output variable the Incidence Rate, which is the Number of Confirmed Cases per 100,000 persons. The third DEA model has an output variable Mortality Rate, which is a Number of Recorded Deaths divided by the Number of Confirmed Cases. The results of this analysis are provided in Table 7.2e below.

Table 7.2e

Module 2 – Average Relative Efficiency Scores.

DEA Model	Group
	High Risk	Mid Risk	Low Risk
Clustering is based on Demographic Risk Factors	Cluster 1	Cluster 2	Cluster 3
(PopDensity, PopDisability, PopPoverty) → Number of Cases	0.86	0.97	0.78
(Physicians Rate, # of Hospitals) → Incidence Rate	0.81	0.83	0.89
(Physicians Rate, # of Hospitals) → Mortality Rate	0.78	0.79	0.89
	High Preval	Mid Preval	LowPreval
Clustering is based on Actual Spread of the Disease	Cluster 3	Cluster 2	Cluster 1
(PopDensity, PopDisability, PopPoverty) → Number of Cases	0.90	0.86	0.88
(Physicians Rate, # of Hospitals) → Incidence Rate	0.86	0.89	0.78
(Physicians Rate, # of Hospitals) → Mortality Rate	0.73	0.89	0.81

The results demonstrate the presence of relative inefficiencies for all three clusters, but we would like to also know whether or not performance of the clusters changed over time, and this is the purpose of the next step.

We used July’s COVID data to run the analysis in the previous step, and in order to investigate changes in the scores of the relative efficiency for each cluster we would have to use the data for multiple time periods. Originally, we considered the following DEA model:

(Population Density, Testing Rate, Incidence Rate) → Mortality Rate.

However, while Testing Rate and Incidence Rate show a low level of correlation with Mortality Rate (0.30 and 0.36, respectively), Testing Rate and Incidence Rate are highly correlated (0.88), thus, we removed Testing Rate from the model. By having data available for April, May, June, July and August we were able to construct 4 time periods (April-May, May-June, June-July, July-August). Additionally, because DEA rewards DMUs with higher levels of outputs per given level of inputs, we needed to convert Mortality Rate in such way, that the states with the lower, and not higher, mortality rates will be rewarded. We converted the output by subtracting the actual reported rate from 10 (the highest original level is 6.10 for Michigan, and the lowest is 1.92 for Tennessee). The summarized results are presented in Table 7.2f below.

Table 7.2f

Module 2 – Average Malmquist Index (MI) Scores.

Change over time (MI)	Group
(Population Density, Incidence Rate) → Mortality Rate	High Risk	Mid Risk	Low Risk
Clustering is based on Demographic Risk Factors	Cluster 1	Cluster 2	Cluster 3
April-May	0.33	0.42	0.47
May-June	1.38	1.38	1.22
June-July	1.32	1.16	1.21
July-August	0.82	0.98	0.85
Average	0.96	0.99	0.94
(Population Density, Incidence Rate) → Mortality Rate	High Preval	Mid Preval	Low Preval
Clustering is based on Actual Spread of the Disease	Cluster 3	Cluster 2	Cluster 1
April-May	0.31	0.41	0.45
May-June	1.49	1.42	1.17
June-July	1.45	1.31	1.05
July-August	0.98	0.81	0.89
Average	1.06	0.99	0.89

It is worth noting that all three risk groups exhibited a significant decline in efficiency (i.e. MI < 1) during the first period of April-May, and for the High Prevalence group the decline was significantly steeper than for the Mid and Low Prevalence groups. However, the consequent three periods have shown a significant improvement in the levels of relative efficiency during May-June and June-July (i.e. MI > 1), followed by a decline in July-August.

It is important to note that a DEA model is not a “true” production model in the sense, let us say, that a recipe is, where (wheat flour, water) → bread. While we can say that flour and water cause bread, and that more flour and more water would cause more bread to be made, we cannot assert, based on DEA alone, that Population Density and Incidence Rate cause Mortality Rate. This is where a decision maker may consider using MR- to investigate if, in fact, the relationships between the inputs and outputs of a DEA model are causal. We illustrate such application of MR to our readers as it applies to our case.

We can assess effectiveness of the states in terms of their fighting the disease by testing the following regression model:

(Population Density, Incidence Rate) → Mortality Rate.

Based on the results of MR (see Table 7.2g ) we can see that during the first three months of the pandemic it is Incidence Rate that impacts Mortality Rate, but during the later period (e.g., July and August) it is Population Density that has a statistically significant impact on Mortality Rate. One of the possible interpretations is that during the beginning of a pandemic the efforts of medical practitioners should be on getting the number of incidents under control (possibly via increase in testing), while during the developed stage of a pandemic the efforts should be allocated towards reducing the density of the population (perhaps by implementing quarantines and social distancing measures).

Table 7.2g

Statistical Analysis of the (Population Density, Incidence Rate) → Mortality Rate link.

Month	R2	Significance F	Variable	Coefficient	P-value
April	0.14	0.04130031	Population Density	−0.0012	0.1484
			Incidence Rate	0.0025	0.0123
May	0.30	0.00183188	Population Density	0.0004	0.7809
			Incidence Rate	0.0023	0.0151
June	0.37	0.000384	Population Density	0.0013	0.4520
			Incidence Rate	0.0023	0.0281
July	0.43	0.0001	Population Density	0.0061	0.0002
			Incidence Rate	0.0000	0.9964
August	0.55	0.0000	Population Density	0.0065	0.0000
			Incidence Rate	−0.0004	0.3406

Also, as we indicated above, Testing Rate and Incidence Rate are highly correlated- this, however, does not imply the presence of causal relationship. But it is the knowledge of the presence of causal relationships that helps a decision maker in fighting the spread of a disease. For example, it is important to pose, and to answer, the following questions:

Does an increase in Testing Rate result in a greater Incidence Rate?

Does an increase in Incidence Rate results in an increase in Mortality Rate?

Consequently, it is of interest to investigate the following sequence of causal links:

Testing Rate → Incidence Rate → Mortality Rate

We do so via 2 stage OLS, where the first model explores the link (see Table 7.2h ), and the second model investigates link (see Table 7.2i ). The reader may observe that for both links, the strength and statistical significance of each link vary depending on the group (or cluster) and the given month. For example, while in April Testing Rate has a statistical significant relationship with Incidence Rate for all clusters and the complete sample, in August no such corresponding relationship existed for Low and Mid Prevalence the clusters though it still existed for the High Prevalence cluster and the complete sample.

Table 7.2h

Statistical Analysis of Testing Rate → Incidence Rate link.

	Module 2 Cluster	R2	Coefficient	P-value
April	Low	0.9610	0.4825	0.0000
	Mid	0.6036	0.3996	0.0049
	High	0.9120	0.2248	0.0001
	Complete Sample	0.7764	0.3908	0.0000
May	Low	0.1749	0.1031	0.1549
	Mid	0.1282	0.0667	0.2530
	High	0.7936	0.3640	0.0029
	Complete Sample	0.7328	0.2923	0.0000
June	Low	0.0700	0.033	0.3821
	Mid	0.1473	0.0599	0.2180
	High	0.8542	0.2087	0.0010
	Complete Sample	0.7318	0.1537	0.0000
July	Low	0.1036	0.0206	0.2834
	Mid	0.0605	−0.0334	0.4408
	High	0.5774	0.0759	0.0286
	Complete Sample	0.5318	0.0809	0.0000
August	Low	0.1357	0.0180	0.2153
	Mid	0.0328	−0.0236	0.5726
	High	0.1477	0.0353	0.3470
	Complete Sample	0.2938	0.0535	0.0011

Table 7.2i

Statistical Analysis of Incidence Rate → Mortality Rate link.

Month	Module 2 Cluster	R2	Coefficient	P-value
April	Low	0.2866	0.0022	0.0594
	Mid	0.0021	0.0002	0.8990
	High	0.0484	0.0020	0.6007
	Complete Sample	0.1321	0.0017	0.0376
May	Low	0.4692	0.0079	0.0097
	Mid	0.0224	0.0012	0.6420
	High	0.4773	0.0020	0.0577
	Complete Sample	0.3413	0.0025	0.0004
June	Low	0.4383	0.0073	0.0136
	Mid	0.3711	0.0032	0.0354
	High	0.6557	0.0027	0.0148
	Complete Sample	0.3966	0.0029	0.0001
July	Low	0.3518	0.0078	0.0326
	Mid	0.0466	0.0008	0.5001
	High	0.3641	0.0045	0.1131
	Complete Sample	0.1471	0.0021	0.0275
August	Low	0.0064	0.0007	0.7942
	Mid	0.0072	0.0001	0.7921
	High	0.0744	−0.0015	0.5133
	Complete Sample	0.0027	0.0001	0.7733

The purpose of this step is to inquire into the sources of the change in the averaged scores of the relative efficiency of the three groups of states. By decomposing the overall change in efficiency (MI) into two of its components, EC and TC, we could gain insights into the drivers of change.

The TC component is associated with the increased availability of a technology. This could be, in the case of our study, increased availability of masks, respirators, gowns, ventilators, and other medical equipment. This also could be associated with the increase in the number of the available hospital beds and medical personnel. Simply put, this component signifies growth driven by the increased availability of the resources. For example, let us consider a scenario of a student who spends 4 h in front of a computer to study for a test to get a grade of 80. If we give this student a better computer (higher resolution, faster processor, more RAM, etc.), and, as a result, the student receives a grade of 85 after studying for test for 4 h, then this growth is driven by change in technology.

The EC component is representative of the improved utilization of the available resources. In the context of our inquiry this could be associated with the implementation of new policies, rules, procedures, and practices according to which the available technology (e.g., personnel, beds, ventilators, etc.) is utilized. By referring to the scenario of the student who studies 4 h to get a grade of 80, if the student receives the grade of 85 all things being equal- study time is 4 h, and the computer is not changed, then this change is due to the EC component.

Ideally, we would like to see the change in efficiency that is balanced- if a student receives a new computer she improves her score because of the better technology AND because she becomes better at using it. In the case of COVID-19, we would like to see the decreased Mortality Rate being due to the increased availability of the needed resources AND due to increased efficiency of utilization of the additional resources. Table 7.2j, Table 7.2k presents data on the EC and TC components and the dominant cause of the change in relative efficiency across pairs of months.

Table 7.2j

Clustering based on Demographic Risk Factors: Comparison of EC vs TC.

	Risk Group
	High			Mid			Low
Period	MI	EC	TC	MI	EC	TC	MI	EC	TC
April-May	0.33	0.60	0.59	0.42	0.60	0.65	0.47	0.74	0.66
May-June	1.38	1.13	1.19	1.38	1.16	1.19	1.22	1.18	1.12
June-July	1.32	0.84	1.58	1.16	0.74	1.53	1.21	0.94	1.30
July-August	0.82	1.30	0.67	0.98	1.40	0.73	0.85	1.00	0.82
Average	0.96	0.97	1.01	0.99	0.98	1.03	0.94	0.97	0.98

Table 7.2k

Clustering based on Actual Spread of the Disease: Comparison of EC vs TC.

	Prevalence Group
	High			Mid			Low
Period	MI	EC	TC	MI	EC	TC	MI	EC	TC
April-May	0.31	0.51	0.63	0.41	0.69	0.58	0.45	0.66	0.67
May-June	1.49	1.19	1.24	1.42	1.21	1.16	1.17	1.02	1.13
June-July	1.45	0.86	1.67	1.31	1.19	1.37	1.05	1.07	1.49
July-August	0.98	1.55	0.64	0.81	1.03	0.80	0.89	1.27	0.72
Average	1.06	1.03	1.05	0.99	1.03	0.98	0.89	1.01	1.00

With respect to the demographic risk factors (see Table 7.2j below), for each of the 3 groups, there were improvements in efficiency (i.e. MI > 1) only in the May-June & June-July periods. For these periods, for the High and Mid Prevalence groups the improvements were primarily due to improved technology (i.e. TC > EC), while for the Low Prevalence group the situation is mixed as for the May-June period the improvement can be attributed to the EC (i.e. EC > TC), while for the June-July period the improvements were primarily due to improved technology (i.e. TC > EC). These results indicate that for the High Prevalence group though for the June-July period there was a decrease in EC, the improvement in TC was sufficient to increase in the corresponding overall efficiency (i.e. MI > 1). However, for the July-August period, while there were improvements in EC component there were not sufficient improvements in the TC component to increase in the corresponding overall efficiency.

With respect to the actual spread of the disease (see Table 7.2k below), for each of the 3 groups, there were improvements in efficiency (i.e. MI > 1) only in the May-June & June-July periods. For these periods, for the High and Low Prevalence groups the improvements were primarily due to improved technology (i.e. TC > EC), while for the Mid threat group the situation is mixed as for the May-June period the improvement can be attributed to the EC (i.e. EC > TC), while for the June-July period the improvements were primarily due to improved technology (TC > EC). Interestingly it is the Mid Prevalence group that shows improvement in EC over 3 of the 4 periods, though sometimes not accompanied by sufficient improvements in TC (i.e. TC < 1). These results indicate that for the High Prevalence group though for the June-July period there was a decrease in EC, the improvement in TC was sufficient to increase in the corresponding overall efficiency (i.e. MI > 1). However, for the July-August period, while there were improvements in EC component there were not sufficient improvements in the TC component to increase in the corresponding overall efficiency. Interestingly the Mid and Low Prevalence groups showed improvements in the EC component in the over each of the last 3 periods.

How could this information, presented in immediate two tables above, be used for a better decision making? Let us consider the guidelines offered by Centers for Disease Control and Prevention for making decisions on the allocation of ventilators to facilities (https://www.cdc.gov/coronavirus/2019-ncov/hcp/ppe-strategy/ventilators.html). The main factors are: 1) Assessment of need; Determination of facilities’ ability to absorb additional ventilators; 2) Ethical considerations to inform how this scarce resource is provided to facilities to save as many lives as possible; 3) Input from state and local leadership, legal and ethical experts, and 4) informed stakeholders. Of especial interest here is the second bullet point- determination of facilities’ ability to absorb additional respirators, where the decision is made based on the following sequence of steps:

Identify facilities that may have capacity to care for critically ill patients who will need mechanical ventilation (from prior or current assessments).

Quantify the number of additional ventilators each facility can realistically absorb.

Base this estimate on having enough trained and qualified staff, space, and necessary equipment needed for caring for additional patients on mechanical ventilation. (Zaza et al., 2016)

Determine the population size that each hospital serves and assess the capacity of each facility to serve vulnerable and high-risk populations within this area.

Consider whether each hospital serves as a referral hospital/regional hospital or serves a high-density population area, rural area, or underserved populations (ibid).

It is easy to see that the decision regarding the estimated quantity of the allocated respirators is based on the available context-specific resources, such as number of trained and qualified staff, space, and supporting equipment. However, as our results suggest, an additional factor to consider is the relative efficiency of the local context to utilize the requested resource- ventilators in this case. It is only expected, that the local context would demonstrate the increase in the TC component of the overall change in efficiency is as a result of the allocation of additional ventilators. However, if we want to obtain a balanced growth in efficiency, then the local context must also exhibit adequate corresponding change in the EC component, which cannot be automatically expected. Instead, it is quite possible (based on the example of the High Prevalence cluster during the June-July period) that the increase via change in technology (i.e. TC > 1) would be corresponded by the decrease via change in efficiency of utilization of the technology (i.e. EC < 1).

The important aspect of the shown above decision making is an implicit assumption of a constant return to scale (CRS). This is because “having enough trained and qualified staff, space, and necessary equipment” presumes, pretty much, a particular ratio that should exist in order to accommodate the currently present, as well as incoming, resource (e.g., ventilators). However, any socio-technical environment, firms, businesses, hospitals, schools, etc. are not perfectly scalable. Thus, we could expect the changes to the ratio specifying the requirements for a particular equipment, not only ventilators.

The DEA-based approach presented above allows for considering not only perfectly scalable constant return to scale (CRS), but also a more suitable variable return to scale (VRS). One of the benefits of such consideration is a more flexible allocation of the needed resources, distribution of which, at this point, has been noted to be uneven (Livingston et al., 2020).

Furthermore, it is worth noting that the capability of assessing the EC and TC components allows for the explicit consideration of the human component of the utilization of resources. This is important, because according to CDC human factors (e.g., respiratory therapists, staff operating the ventilators) serve as a bottleneck (https://www.nationalgeographic.com/science/2020/03/us-america-has-fraction-medical-supplies-it-needs-to-combat-coronavirus/) in applying physical resources to treatment of a disease. If we apply our DSS at the hospital level, then, knowing that there are approximately 20 technicians capable of operating ventilators per hospital (https://www.bls.gov/ooh/healthcare/respiratory-therapists.htm), we should be able to obtain a context-sensitive representation of the relative efficiency of utilization of supplies and equipment allowing for more appropriate allocation of what is needed.

Conclusion

Delivery of a modern healthcare is an increasingly multidimensional undertaking that requires optimization of the provision of health services, be it in the context of emergency departments (Cabrera et al., 2011), or prescription of medicine (Sintchenko et al., 2008), or managing a Covid-19 pandemic (Mora et al., 2021). This increase in dimensionality of the problem results in the evolution of DSS used in healthcare in terms of their complexity (Safwan et al., 2016), thus requiring their reliance on increasingly more complex components (e.g., data mining and machine learning) (Shailaja et al., 2018). At this point, it has been reported that applications of DSS to optimize provision of healthcare in the context of Covid-19 was very limited and without incorporating such fundamental aspects as feasibility and health system considerations and consideration of the advanced methods of data analysis (Mora et al., 2021). For example, there were efforts to contribute via creating DSS targeting physical distancing (Adam et al., 2021) and available food supply (Blackmon et al., 2021). And while it might be possible that DSS developed for a different purpose and a different industry (e.g., bankruptcy prediction) could be adapted to address some pandemic related issues (Perboli & Arabnezhad, 2021), it can also be advantageous to develop a target-specific healthcare DSS (Sutton et al., 2020) for dealing with the spread of the infectious disease. And this is the route we followed in our work.

Despite the emergence of DSS as one of the premier exemplars of IT in healthcare, it is primarily due to the complexity of the systems that their use is not wide spread (Rajalakshmi et al., 2011, Wasylewicz and Scheepers-Hoeks, 2019). It is difficult to reconcile the desire for simplicity of DSS in healthcare with the call for the increase in their functionality, where the normal targets (e.g., quality, risk, productivity, etc.) are to be supplemented with pattern recognition and proactive decision making (Kohli & Piontek, 2008), but this is exactly what we attempted to do in this investigation, for the benefits are worth the effort (Latif et al., 2020).

A crucial component to a success of a DSS in fighting a pandemic is the data and their source, in our case reflecting Covid-19 data (Guidotti and Ardia, 2020, Wang et al., 2020, Chen et al., 2020). There are multiple global data repositories available (e.g., IHME (http://www.healthdata.org/covid/data-downloads), LANL-GR (https://covid-19.bsvgateway.org/), USC SIKJalpha (https://github.com/scc-usc/ReCOVER-COVID-19), Johns Hopkins University Center for Systems Science and Engineering (JHU CSSE) (https://github.com/CSSEGISandData/COVID-19), etc.) but the main focus of several investigators using the data was on forecasts and alternative scenarios of Covid-19 mortality as being critical inputs in fighting the pandemic (Friedman et al., 2021). Khan et al. (2021) noted that the applications of AI in the fight against Covid-19 could be categorized as focusing on diagnosis, screening, prediction, and drug repurposing. However, none of previous Covid-19 research explicitly considered the issue of efficiency.

In this paper we presented and illustrated a conceptual model for an easy-to-construct modular DSS that would be useful for addressing a variety of public health questions related to the occurrence of a ‘new’ disease. Our research can be considered to be complementary to the previous Covid-19 studies, for it deals with identifying a preferential route of the spread of the disease, and with the assessment of the efficiency of utilization of the available healthcare resources once the disease penetrated a new location. Benefits of using our DSS conceptual model include:

Provides for discovering naturally occurring groups based on the demographic risk factors

Provides for identifying the sources of heterogeneity between the groups

For each time period, provides for:

Estimating the relative efficiency of each group with for example the disease containment

Determining whether there are improvements in efficiency

Identifying whether the dominant cause (better technology, or better processes including better utilization of technology) for changes in relative efficiency in each group

Provides for uncovering sample and group-specific non-obvious causal structures

Identifying causal impact of the risk factors on the contraction of the disease

Identification of High, Mid, and Low risk naturally occurring groups based on the efficiency and effectiveness in containing the spread of the diseases (and minimizing the mortality rate) with a correspondent suggestion for the improvements.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.