Picture fuzzy set-based decision-making approach using Dempster-Shafer theory of evidence and grey relation analysis and its application in COVID-19 medicine selection.

To offer better treatment for a COVID-19 patient, preferable medicine selection has become a challenging task for most of the medical practitioners as there is no such proven information regarding it. This article proposes a decision-making approach for preferable medicine selection using picture fuzzy set (PFS), Dempster-Shafer (D-S) theory of evidence and grey relational analysis (GRA). PFS is an extended version of the intuitionistic fuzzy set, where in addition to membership and non-membership grade, neutral and refusal membership grades are used to solve uncertain real-life problems more efficiently. Hence, we attempt to use it in this article to solve the mentioned problem. Previously, researchers considered the neutral membership grade of the PFS similar to the other two membership values (positive and negative) as applied to the decision-making method. In this study, we explore that neutral membership grade can be associated with probabilistic uncertainty which is measured using D-S theory of evidence and FUSH operation is applied for the aggregation purpose. Then GRA is used to measure the performance among the set of parameters which are in conflict and contradiction with each other. In this process, we propose an alternative group decision-making approach by the evidence of the neutral membership grade which is measured by the D-S theory and the conflict and contradiction among the criteria are managed by GRA. Finally, the proposed approach is demonstrated to solve the COVID-19 medicine selection problem.

Introduction

Traditional logic, which is interpreted as either true or false, found to be difficult to solve uncertain real-life problems. As a counter measure, Zadeh (1965) invented fuzzy set theory, where the involvement of elements in a set is characterized by membership grade, which belongs to [0, 1]. To handle much uncertainty, fuzzy sets were extended by the different researchers in different ways such as vague set (Gau and Buehrer 1993), intuitionistic fuzzy set (IFS) (Atanassov 1986a, 1986b), fuzzy soft set (Das et al. 2018), rough set (Pawlak 1982), fuzzy interval theory (Gorzalczany 1987), intuitionistic multi fuzzy set (Das et al. 2013), interval-valued intuitionistic fuzzy set (Park et al. 2008), intuitionistic fuzzy soft set (Deng 1982) and neutrosophic soft set (Das et al. 2019). Consequently, the application of fuzzy set theory and its extensions increased rapidly in the decision-making methods in various domains like medical diagnosis (Das et al. 2013), pattern recognition (Wei and Lan 2008), data analysis (Zou and Xiao 2008), forecasting (Xiao et al. 2011), optimization (Kovkov et al. 2007), simulation (Kalayathankal and Singh 2010) and texture classification (Mushrif et al. 2006). Recently in 2014, Cuong (2014) developed the picture fuzzy set (PFS) as the generalized form of fuzzy set and IFS. The PFS approaches are found to be more appropriate in those cases when the views of someone contain more option types like yes, abstain, no and refusal. The general election of a country is noted as a good example to describe PFS, where a voter can cast his vote in favour of the candidate (yes), against the candidate (no), may not cast his vote (abstain) or may refuse to cast his vote in favour of the given candidates and prefer for nota (refusal) (Cong and Son 2015). From the time of its introduction in 2014, many researchers have been contributing to the development of decision-making problems using PFS. A number of multi-criteria decision-making (MCDM) approaches have been developed to manage real-life problems in the domain of picture fuzzy sets. By studying the intuitionistic fuzzy aggregation operator, Wang et al. (2017) described some picture fuzzy geometric operators and generalized the basic properties of those operators. Then they applied the proposed operators in multiple attribute decision-making problems under the picture fuzzy domain. Si et al. (2019) introduced a novel ranking method to rank the picture fuzzy numbers (PFNs). Initially, they compared the PFNs, even when the accuracy and score values of those PFNs are equal. In Si et al. (2019), the ranking method is based on positive ideal solution, positive and negative goal differences, and score and accuracy degrees of the PFNs. Another score function was developed to estimate the actual score value that depends on the positive and negative goal differences and the neutral degree. Garg et al. (2017) proposed a sequence of aggregation operators, namely picture fuzzy weighted average aggregation operators, picture fuzzy hybrid average aggregation operators and picture fuzzy ordered weighted average aggregation operators and used them in decision-making problems. Wei et al. (2016) extended the cross-entropy of fuzzy sets in the context of PFS and developed the picture fuzzy cross-entropy to solve the multiple attributes decision-making (MADM) problem. Wei et al. (2018a) expanded the TODIM model with the picture fuzzy members (PFNs) and generated the relative weight of all attributes and calculated dominance degree of alternative in respect to all other alternatives to develop dominance matrix. Then they computed the overall dominance degree of each alternative and determined the alternatives’ ranking based on it. Son et al. (2017) introduced and extended the fundamental distance measure using picture fuzzy sets and proposed generalized picture distance measures and picture fuzzy association measures. Son et al. (2016) developed a picture fuzzy set-based distance measurement technique and applied it in the picture fuzzy-based clustering methods. The propositions of various similarity measures such as cosine similarity measure and weighted cosine similarity measure between PFSs were studied by Wei et al. (2018b). Then, the authors applied the similarity measurement methods to detect the building material and recognize the minerals field. Most of the researchers, who investigated in the picture fuzzy domain, considered the neutral membership degree similar to the other two degrees of positive and negative membership during the evaluation of the decision-making problem. But sometimes, it is found to be difficult to differentiate the neutral membership grade from the positive and negative membership grades.

Fuzzy sets and extended fuzzy sets are well used to manage uncertainty and vagueness. Besides these, there are other types of uncertainty induced by random phenomena that are called probabilistic uncertainty. But the probabilistic uncertainty is not found to be enough to consider the various uncertain evidence. Dempster (1968) introduced the belief function that presented the subjective assessments by using probability. Then Shafer (1976) extended the Dempster concept into a mathematical theory of evidence where it remains a classic in belief function. That is why this is called the theory of evidence. The combined concept of both of them is called the Dempster–Shafer (D–S) theory of evidence. The theory of belief functions or the D–S theory (Shafer 1976) is a mathematical framework of evidence that can be deduced as a generalization of probability theory. According to the D–S theory, the incidents belong to the sample region to which the nonzero probability mass of the attributes is not a single point but sets. In (Beynon et al. 2000), the authors introduced the basic concepts of the D–S theory of evidence regarding probability and compared it with the traditional Bayesian theory. Pankratova and Nedashkovskaya (2013) presented a mathematical analysis of the sensitivity of diverse combination rules hybridizing D–S theory and analytical hierarchy process (AHP) to solve foresight problems. Donga and Xiao (2015) introduced the hybrid concept of the Dempster–Shafer fuzzy soft sets through the combination of Dempster–Shafer theory and fuzzy soft sets. They developed the FUSE operator applied on Dempster–Shafer fuzzy soft sets and created the relationship between incomplete fuzzy soft sets and D–S fuzzy soft sets. Dutta and Ali (2011) discussed the Dempster–Shafer theory of evidence by considering focal elements as triangular fuzzy numbers. Then they formulated a method for obtaining belief and plausibility measure from the basic probability assignments (BPAs) assigned to fuzzy foal elements. Finally, they used the fuzzy focal elements in Dempster–Shafer theory and executed it to evaluate the human health risk (non-cancer) evaluation process with hypothetical data. The sets that get nonzero mass are considered as the focal elements. The summation of these probability masses is one; however, the basic difference between D–S theory of evidence and conventional probability theory is that the focal elements of a D–S formation may overlap one another. The D–S theory of evidence also provides a way to symbolize and merge weights of evidence.

Grey relational analysis (GRA) is a significant part of grey system (GS) theory introduced by Deng in 1982 (Deng 1982). GS presents a system where some part of the information is known and the remaining part of information is unknown. According to this definition, some part of the information is qualitative and quantitative from the entire information. Due to lack of information, some information belongs to the grey area. The uncertainty may be present in the different position within the entire region or may be within the grey area. GRA is more effective to solve the problem where complicated relationships are noticeable among the factors and variables. GRA (Wei and Lan 2008; Deng 1982) provides a collection of strong sets of statements about system solutions. The system is unable to provide any solution if the system does not have any information. Alternatively, the systems generate a unique and acceptable solution due to the availability of perfect information. In the presence of either complete or partial information, grey systems will give a variety of available solutions. Most of the developed picture fizzy number-based MCDM methods calculate the distance between two PFNs. But those methods are unable to estimate the individual importance of the membership values. Suppose one pair of PFNs are and , and another pair of PFNs are and , where , , and , respectively, denotes positive membership, neutral and negative membership grade of PFN. Here we consider . Thereafter, the distance of each pair of PFNs and is same, yet one is unable to realize the importance of higher membership values , of the first pair, i.e.. Those membership degrees have individual importance in different directions.

The most acceptable ideal situation has the maximum value of positive membership degree, a minimum value of negative membership degree and average value neutral member degree. Maximum values of positive membership degree are highly desirable to fulfil most of the criteria, and accordingly, minimum value of negative membership degree is less desirable. Whenever the neutral member degrees are found the same for all the sets, then the neutral member degrees have no effect on decision-making. In these circumstances, the most suitable procedure is GRA, which has three individual estimations for maximization, minimization and averaging of positive, negative and neutral membership degrees. The better solutions will be the larger, smaller and average, respectively. GRA is capable of managing the complex connection between parameters. Grey analysis does not attempt to find the best solution but provide a technique for determining a good solution suitable to solve real-world problems (Kuo et al. 2008). This theory motivates the researchers for generating acceptable solutions in grey scenarios and then to upgrade it in a number of ways.

Nowadays, the whole world has become fully unbalanced and passing through an uncontrolled situation due to the dangerous and novel virus COVID-19. Most countries are totally stagnant and the people are quarantined to make themselves safe from COVID-19 (Ren et al. 2020). Many researchers are continuously contributing to developing various type of mathematical and hybrid models to predict the future trends, strength and transmission capability of COVID-19 virus, and have drawn some useful conclusions which assist the health department to take the necessary precaution to track and handle the COVID-19 situations. The authors in Melin et al. (2020) introduced a novel hybrid prediction model that can mergethe ensemble architectures of fuzzy logic-based neural networks for response integration. The fundamental concept of the proposed model is to merge several fuzzy-based neural network predictors, control the uncertainty of the individual networks and try to reduce the uncertainty of the total predictions. This model was able to predict the future trends of COVID-19 up to some extent and help the authorities make the necessary decision to handle the health care system in a better manner. The authors in Sun and Wang (2020) collected the COVID-19 data from a decided location within a specific time interval and trained through the ordinary differential equation model for fitting. Then, they modified the simulation by the trained model to realize the effect of the COVID-19 affected visitors. They found that the affected visitors have a great role in the newly introduced case of COVID-19. Stochastic simulations proved that the physical connections could be rapidly increased due to the affected visitors which are considered sufficient for the local outbreak of COVID-19. The confirmed case of asymptomatic patients was significantly less than the model predictions quantity. This indicated that a major portion of asymptomatic patients are not identified/found. Fuzzy-based hybrid approaches for forecasting the confirmed cases and deaths of the countries according to their time series are given in Castillo and Melin (2020). The fundamental concept of this proposed hybrid method (Castillo and Melin 2020) is to combine the fractal dimension and fuzzy logic for enabling efficient and accurate forecasting of COVID-19 time series. The fractal dimension is provided to differentiate and categorize the object. They introduced a fuzzy rule-based system to represent the knowledge about the forecasting time series of the countries. The authors in Castillo and Melin (2021) introduced the hybrid procedure for composing the fuzzy logic and fractal dimension which measured the uncommon activities of times series to classify countries according to their COVID-19 time series data. The proposed method generates an accurate classification of countries based on the complexity of the COVID-19 time series data. Editors (Boccaletti et al. 2020) of the journal “Chaos, Solitons and Fractals” analysed the impact of COVID-19 pandemic throughout the world and felt the necessity to create a unique platform for the researchers to help the society to avoid the worst effects of future pandemics. Recently, Mishra et al. (2021) proposed an extended fuzzy decision-making framework using hesitant fuzzy sets for the drug selection to treat the mild symptoms of COVID-19. Although the researchers are working hard, they are still struggling to recover from this unwanted situation. The scientists from different domains are consistently trying to apply their knowledge in different perspectives such as dominating the virus, identifying the virus, isolating from the virus, protecting from the virus, and finding the treatment of the virus affected patients, to manage the superfluous situation (Kumar et al. , 2020; Ghosh et al. , 2020), which are considered to be the long-term project. As an intermediate solution, the most important aspect is to provide suitable medical service to the affected patients and recover those who are critically ill due to perilous virus COVID-19. The health department of India has classified the COVID-19 affected patients into some categories according to the patient's physical condition. The extreme condition is called severe cases, and this type of patient requires quality treatment (Clinical Management Protocol 2020). The health workers provide some probable treatment (Clinical Management Protocol 2020) to cure the unpredicted virus infection due to the non-existence of any kinds of approved treatment, where the selection of medicines has a huge impact on the recovery rate of the patients. As found in the literature, a few researchers have experimented on the selection of medicines for the COVID-19 affected patients. The proposed medicines for treating COVID-19 affected patients have various functionalities like effectiveness, side effect and some unseen effects that are uncertain.

To fill up this research gap, this paper proposes an alternative PFS-based approach using the group multi-criteria decision-making problem to explore the suitable medicines that are considered the most urgent to save the lives of the affected persons. In PFS, we find that among the four membership degrees (positive, negative, neutral and refusal), the neutral membership degree is fully unpredictable and undecided as the positives or negatives degrees of it are completely unknown. But the PFS is needed to be used in managing some real-life situations so that the experts can express their observations and judgment in the form of PFSs. In this paper, D–S theory is applied to the PFS framework to estimate the evidence of the neutral part. In this study, we apply FUSH operation to merge the opinions recommended by the experts in the form of picture fuzzy numbers. In the process, resultant PFS is formed to incorporate the opinions of multiple experts. According to the grey relational analysis, the grey relational grades are evaluated according to grey relational coefficient of the alternatives. We use the proposed approach to select suitable medicines for the affected persons in the context of PFS.

The rest of the article is arranged subsequently. In Sect. 2, we discuss PFS. Then, in Sect. 3, we describe the basic concept of D–S theory and discuss some modified concepts of D–S theory and FUSH operator using PFS. Next, a detailed discussion on grey-based relational analysis is given in Sect. 4. We have presented a PFS-based group decision-making method using the D–S theory and GRA in Sect. 5. A numerical example of the proposed method is stepwise discussed in Sect. 6. Then the proposed method is evaluated for COVID-19 medicine selection in Sect. 7. In Sect. 8, we compare the projected method with three existing methods with respect to some conflicting situations. Then we verify the validity of the projected method by the three generalized criteria in Sect. 9. Lastly, the key observations are drawn in Sect. 10.

Picture fuzzy sets

A picture fuzzy set on the universe X is defined aswhere be the degree of positive membership of in P, similarly and are, respectively, called the degrees of neutral and negative membership of in P. These three parameters of the picture fuzzy set P satisfy the following condition .

Then, the refusal membership grade of in P can be calculated in the following way,

The neutral membership of in P can be thought as degree of positive membership as well as degree of negative membership, whereas refusal membership can be explained as not to take care of the system. When , then the PFS reduces into IFS.

For a fixed is defined as a picture fuzzy number (PFN), where ,,, and

Simply, PFN is signified as .

Operations on PFS

For two PFSs and , Cong (2014) defined some operations as given below.

Cuong and Kreinovich (2015) and Cuong (2017) defined some properties on PFSs as given below.

If and then ;

;

Distance between PFSs

Distances between the two PFSs are defined in Cuong and Kreinovich (2014); Cong and Son 2015; Si et al. 2019). The distance between two PFSs and in is calculated as given below.

Normalized Hamming distance

Normalized Euclidean distance

Example 1

Let P = {(0.7, 0.2, 0.1), (0.8, 0.1, 0.1), (0.7, 0.1, 0.2)} and N = {(0.6, 0.2, 0.2), (0.8, 0.2, 0.0), (0.9, 0.0, 0.1)} are two picture fuzzy sets of dimensions 3. Then.

Wang et al. (2017) defined some special operations of the picture fuzzy set. They proposed the following operations on PFNs and .

Example 2

Let P = (0.7, 0.2, 0.1) and N = (0.6, 0.2, 0.2) are two PFSs and λ = 5.

Comparison of picture fuzzy sets

Wang et al. (2017) used the accuracy function and score function to compare the PFSs. Let be a picture fuzzy number (PFN), then a score function is defined as and the accuracy function is given by where and . Then, for two PFNs M and T.

If , then M is higher than T, denoted by M > T;

If , then

, implies that M is equivalent to T, denoted by M = T

, implies that M is higher than T, denoted by M > T.

Example 3

Let M = (0.7, 0.2, 0.1) and T = (0.6, 0.2, 0.2) are two picture fuzzy sets. Now,

= 0.7–0.1 = 0.6,  = 0.6–0.2 = 0.4

= 0.7 + 0.2 + 0.1 = 0.9,  = 0.6 + 0.2 + 0.2 = 1.

Since , therefore M > T.

D–S theory of evidence

D–S theory is a mathematical concept of combining evidences based on belief functions and plausibility reasoning. It is used to combine the evidences in order to compute the probability of an event and uses the idea of “mass” as opposed to Bayes theory. It is known as the theory of evidence because it handles the weight of evidence. In order to measure the uncertainty of an event, D–S theory applies an interval [belief, plausibility] where belief is a measure of the strength of evidence in support of a subset and it represents the evidence we have for it directly, whereas plausibility represents the maximum share of the evidence we could possibly have. Dempster–Shafer theory (Shafer 1976) states our assumption between a universe of discourse U and a set of corresponding statements to a group of propositions, where only one statement is true. The propositions are assumed to be complete, mutually exclusive and exhaustive. Let 2 denote all subsets of U including itself and empty set, so the total numbers of subsets are 2. The basic probability assignment (bpa) function on 2 is defined below.

Consider is a proposition, is the evidence support of and is called the degree of ignorance, where is called focal elements of for every subset . In PFS ,, and are considered as focal elements which, respectively, denote the positive membership, neutral membership and negativemembership value of the alternatives and criteria of the PFS .

The belief function can be defined as which is mapped into [0, 1] and is computed as

Another function called plausibility function of , denoted as , is defined as follows:

The total belief of is represented by , whereas measure the total belief that arises under . Then the and are called lower bound function and upper bound function, respectively, denoted as . The relation between and is defined as follows:

The uncertainty of the object can be represented as:

The facts described above are illustrated using the example given in Table 1 which shows a combination of concordant evidence using D–S theory.

Table 1

Example of belief and mass function

Function	{β}	{γ}	{δ}	{β,γ}	{γ,δ}	{β,δ}	{β,γ,δ}
pf	0.05	0.1	0.05	0.1	0.1	0.1	0.5
Bel	0.05	0.1	0.05	0.25	0.25	0.2	1
Pl	0.75	0.8	0.75	0.95	0.95	0.9	1
u	0.7	0.7	0.7	0.7	0.7	0.7	0

Suppose that a patient suffers with three symptoms like high fever (β), dry cough (γ) and tiredness (δ), and then the patient is suspected to be affected by COVID-19. Hence, the frame of discernment is represented by  = {β, γ, δ}. We have considered that the evidence (m) of COVID-19 has been collected and represented by the basic probability assignment (). The symbols used in Table 1 represent the above-mentioned fact.

Dempster's rule of combination

Based on D–S theory, the evidence, belief, plausibility and uncertainty values of a single tone event by basic probability assignment () function are measured above in Sect. 3. Subsequently, when two or more events are present, the situation can be managed by Dempster's rule of combination as mentioned below. Multiple belief functions in Dempster’s rule are combined using their basic probability assignments (m).

Consider the two evidences and of the power set ϴ for which the corresponding respective basic probability assignment functions are and . The evidence combination rule for the two different evidences and is defined below, where be the combined evidence of and .

Here indicates the conflict combination of the evidence named as conflict percentage. The combined evidence needs to be normalized with respect to other combinations, which depends on the normalized factor .

Example 4

Suppose that two teachers are appointed to evaluate 100 students of a particular class for assessment. As per evaluation of teacher 1, 50 students score grade X and 30 students’ score grade Y. According to evaluation of teacher 2, 40 students score grade X and 40 students score either X or Y. Then, the evidence regarding the students' scores are combined and the resultant evidence is determined. Here, is considered as a frame of discernment and be the power set. We have considered m1 and m2 as mass functions corresponding to teacher 1 and teacher 2. The combination of concordant evidence using D–S theory is shown in Table 2.

Table 2

Combination of concordant evidence using D–S theory

Evidence	pf_1(X) = 0.5	pf_1(Y) = 0.3	pf_1(ω) = 0.2
pf_2(X) = 0.4	pf_12(X) = 0.2	pf_12(Θ) = 0.12	pf_12(X) = 0.08
pf_2(X,Y) = 0.4	pf_12(X) = 0.2	pf_12(Y) = 0.12	pf_12(X,Y) = 0.08
pf_2(ω) = 0.2	pf_12(X) = 0.1	pf_12(Y) = 0.06	pf_12(ω) = 0.04

According to the expression (2),  = 0.12. Then according to expression (1), we are getting the evidence of X and Y as given below.

FUSH operator

In order to perform the proposed FUSH operation, the datasets in terms of PFNs are collected from different reliable sources, and then those datasets are combined and new PFSs is generated which is more reliable and authentic than the inputs. This process is called data fusion and the operator used to combine the data sets is called the FUSH operator which is defined below:

Here and are the two input PFSs and we obtain the resultant PFS using three FUSH operations for three parameters of PFS as shown below. In this study, the evidences concerned with positive and negative membership grades are completely given, whereas the evidence corresponding to the neutral membership grade is associated with the positive and negative membership grades. The expression (3) and expression (5) consider the evidence of any propositions of and , respectively, whereas expression (4) explicitly depends on .

Example 5

Let P = (0.7, 0.2, 0.1) and N = (0.6, 0.2, 0.2) are two PFNs and and are the mass functions regarding the membership functions (positive, neutral and negative) of picture fuzzy numbers P and N (Table 3).

Table 3

Combination of concordant evidence using D–S theory and PFN

Evidences	pf_1(μ) = 0.7	pf_1(η) = 0.2	pf_1(ν) = 0.1
pf_2(μ) = 0.6	pf_12(μ) = 0.42	pf_12(ημ) = 0.12	pf_12(Θ) = 0.06
pf_2(η) = 0.2	pf_12(ημ) = 0.14	pf_12(η) = 0.04	pf_12(ην) = 0.02
pf_2(ν) = 0.2	pf_12(Θ) = 0.14	pf_12(ην) = 0.04	pf_12(ν) = 0.02

We obtain , and according to expressions (3), (4), and (5), respectively, which are shown below.

Then the resultant PFN after FUSH operation between PFN and is obtained as = (0.68, 0.04, 0.08).

Note that FUSH operator follows the associate properties that indicate the three PFNs , and hold the relation defined below.

GRA

In MCDM method, the criteria values of the alternatives are of different units and have different influence in the decision-making process due to different ranges of the criteria values (Deng 1982; Kuo et al. 2008). Due to this difference in the units and large interval of criteria values, often incorrect results are generated during the analysis. Hence normalization of all the performance values for all criteria is unavoidable in order to convert them into a comparable sequence. In order to analyse and utilize the GRA model, consider a MCDM problem with m number of alternatives and n number of criteria, where the alternative is represented as . Here indicates the observation value of the criteria of alternative. The information regarding alternative is converted into comparable sequence of using one of the following three equations defined below.

Here the symbols and represent the maximum and minimum value of , respectively, and be the pivot value which is chosen by the expert arbitrarily depending on the problem type. Expression (6) is used to normalize the maximum value as optimum one behaviour type criteria, expression (7) is used to normalize the minimum value as the optimum one behaviour type criteria, whereas expression (8) is used for normalize the criteria which is nearest to pivot value and considered as optimum. In the proposed study, the decision-making method considers the expressions (6), (7), and (8) for normalizing the grey relation of the positive, negative and neutral membership grades of the criteria, respectively.

After normalizing all observation values between 0 and 1, the observation value of the criteria with respect to the alternative is considered as . Then the support of criteria depends on the value of . If the value of is 1 or closer to 1, then the criteria are most acceptable for the alternative in respect to all other criteria. An alternative is highly acceptable when all the observation values are 1 or near to 1. But this type of ideal situation is not generally found. The ideal solution is represented by a reference sequence of combination of 0 and 1, where the reference sequence is denoted as . Then, an alternative is searched which is closer to the reference sequence. The closeness between an alternative and the reference sequence is measured by the parameter which is known as the grey relation coefficient. The grey relational coefficient represents the relationship between the experimental result of ideal and normalized information. The grey relation coefficient is denoted as and calculated by the following equation.

Here , and is the distinguishing coefficient,. The distinguishing coefficient controls the range of the grey relational coefficient and efficient value of the coefficient is 0.5. Let us consider three alternative c, d and e, and and for the criteria , where alternative d is found to be the most nearest to the reference sequence. Generally, the values of and are, respectively, closer to 1 and 0. The grey relational grades of the alternatives are calculated by the addition of all grey relational coefficients of the respective alternative as defined in (10).

Here is represented as the weight of the criteria and the distribution of among the criteria is dependent on the expert’s opinions. The degree of similarity depends on the value of grey relational grade. The highest value of grey relational grade is the closer to the reference sequence and acceptable option.

Proposed decision-making approach based on PFS

Here we develop a new evidence-based logical decision-making method using PFN for solving the MCDM problem, where PFN is used to represent the decision information provided by the decision makers. Let be the finite set of alternatives, be the set of criteria, and be the decision matrix in the form of PFNs. The notation and represent the positive membership, neutral membership and negative membership degree of alternative in respect of criteria. The proposed algorithmic approach is given below in stepwise manner.

Step 1 We present the evaluating values of alternative corresponding to the criteria using PFN

Step 2 Merge all the criteria information of an individual alternative into a collective PFN using the FUSH operation which is defined in Sect. 3.2.

Step 3 Compute the comparable sequence for each of the alternatives from the collective PFNs , where , , and are, respectively, obtained using expressions defined in (6), (8) and (7).

Step 4 Calculate the grey relational coefficient for each alternative to measure the closeness between the comparable sequences of alternatives and reference sequence using (9).

Step 5 The membershipwise (positive, neutral and negative) grey relational grade for each alternative is computed based on for computing the actual grey relational grade using expression (10).

Step 6 Determine the ranking order of the alternatives according to the calculated actual grey relational grade , where higher value of indicates better rank.

Numerical analysis

This section presents a numerical example to illustrate the proposed approach. We consider five alternatives and five criteria , where the evaluating values of the criteria regarding the alternatives are expressed using PFNs. Those criteria are neither fully supported nor fully rejected by the set of alternatives due to the existence of neutral membership parameters. Thereafter, the decision matrix M is presented in the form of picture fuzzy information, which is shown in Table 4.

Table 4

Decision matrix M

	C1	C2	C3	C4	C5
A1	[0.21,0.64, 0.14]	[0.39, 0.44,0.17]	[0.21, 0.29, 0.5]	[0.07,0.47, 0.47]	[0.12, 0.31, 0.56]
A2	[0.17, 0.33, 0.5]	[0.37, 0.26, 0.37]	[0.42, 0.05, 0.53]	[0.2, 0.4, 0.4]	[0.53, 0.18, 0.29]
A3	[0.4, 0.2, 0.4]	[0.45, 0.2, 0.35]	[0.06, 0.56, 0.38]	[0.18, 0.59, 0.24]	[0.6, 0.1, 0.3]
A4	[0.6, 0.2, 0.2]	[0.21, 0.42, 0.38]	[0.47, 0.37, 0.16]	[0.3, 0.43, 0.26]	[0.8, 0.1, 0.1]
A5	[0.6, 0.13, 0.27]	[0.27, 0.41, 0.32]	[0.31, 0.15, 0.54]	[0.29, 0.41,0.29]	[0.42, 0.25, 0.33]

According to the FUSH operation defined in Sect. 3.2, we perform the FUSH operation among the PFNs of various the criteria corresponding to each alternative and generate the collective decision matrix which is displayed in Table 5.

Table 5

Collective decision matrix

Alternative	Collective grade
A1	[0.07, 0.01, 0.3]
A2	[0.06, 0, 0.11]
A3	[0.13, 0, 0.1]
A4	[0.28, 0, 0.02]
A5	[0.11, 0, 0.08]

Next the comparability sequence is generated from the collective decision matrix by the grey relational generating process which is defined in (6), (7) and (8). Generated compatibility sequence is presented in Table 6.

Table 6

Comparability sequence of the collective decision matrix D

Alternative	Compatibility sequence
A1	[0.05, 0.98, 0]
A2	[0, 1, 0.68]
A3	[0.32, 1, 0.71]
A4	[1, 1, 1]
A5	[0.23, 1, 0.79]

Then we calculate the grey relational coefficient based on membership degree of the alternatives in respect to the criterion. This is done using the expression (9) based on the compatibility sequence. The resultant grey relational coefficient is shown in Table 7. During the execution, we consider the distinguished coefficient value as 0.5.

Table 7

Grey relational coefficient

Alternative	Grey relational coefficient
A1	[0.95, 0.02, 1.0]
A2	[1.00, 0.0, 0.32]
A3	[0.68, 0.0, 0.29]
A4	[0, 0, 0]
A5	[0.77, 0, 0.21]

Next the grey relation grades of the various membership degrees (positive, neutral and negative) are computed. Finally, the actual grey relation grades of the alternatives are obtained by adding the grey relation coefficient of the respective membership degrees using expression (10). Here, the weights of all membership degrees are considered to be equal. The resultant actual grey relation grades are displayed in Table 8.

Table 8

Grey relational grade of the alternatives

Alternative	Grey relational grade	Actual grey grade
A1	[0.51, 0.02,0.50]	1.03
A2	[0.50, 0.02, 0.76]	1.28
A3	[0.59, 0.02, 0.78]	1.39
A4	[1.0, 0.02, 1.0]	2.02
A5	[0.56, 0. 02, 0.82]	1.41

Finally, the alternatives are ranked according to the grey relational grade.

Therefore, the order of the alternatives is as follows: .

COVID-19 medicine selection

This section presents and analyses the usefulness of the proposed technique. As per the information available and our knowledge, there are only a few approved medicines for the treatment of corona positive patients. Henceforth in this paper, we consider the investigational therapies for the COVID-19 patients, where there are many medicines and some of the medicines are considered better and acceptable. COVID-19 is an encrypted form of a large group of viruses (coronavirus) that infects the human and animal bodies that are the main causes for illness. This virus has the enormous potential to be spread from the infected people or animals to other people either by physical contact or through small air transmissions like MERS and SARS virus. As per the information available, the outbreak of novel coronavirus (COVID-19) was noted from Wuhan city in Hubei of the Republic of China in the month of December, 2019 (Ren et al. 2020). At present, most countries around the world (214 countries) are facing challenging problems due to COVID-19 infection. By observing the dangerous impact of coronavirus on the human population, the International Health Regulations (WHO) has already declared this outbreak as a “Public Health Emergency of International Concern” (PHEIC) on 30th January, 2020 and marked it as pandemic on March 11, 2020 (Clinical Management Protocol 2020). Medical experts have observed that COVID-19 patients mainly suffer from the following symptoms and signs: fever, fatigue, cough, shortness of breath, myalgia, expectoration, rhinorrhea, diarrhoea, sore throat, loss of taste (agues) or loss of smell (anosmia), and in severe cases respiratory symptoms have also been reported. Older people and particularly immune-suppressed patients often feel typical symptoms such as fatigue, reduced mobility, reduced alertness, diarrhoea, delirium, loss of appetite and absence of fever. As per information available from the Integrated Health Information Platform (IHIP)/Integrated Disease Surveillance Programme (IDSP), Govt. of India, dated 11.06.2020, 15,366 corona positive samples were collected for experimentation. Among those 15,366 samples, the signs and symptoms for running nose, weakness, breathlessness, sore throat, cough, fever, and others were, respectively, 3%, 7%, 8%, 10%, 21%, 27% and 24% (Clinical Management Protocol 2020). As per the Ministry of Human and Family Welfare, Govt of India notification (as on 19th November, 2020), around 10 million people are affected. Among the affected people, 9.2 million are recovered, 0.45 million are active cases and 1.3 million have lost their life. The health department is consistently trying to provide better treatment to the COVID-19 affected people using the past experience of controlling the pandemic events.

Unfortunately, no specific treatments have been approved for the COVID-19 affected people. Several therapies or approaches are considered for managing or curing the COVID-19 patients. The symptomatic treatment of the COVID-19 patients is provided in different ways like mild cases, moderate cases and severs cases. Presently, on the basis of the limited available evidence, health experts use these therapies. Depending on the situation and availability of relevant data, the evidence can be incorporated, and recommendations can be upgraded accordingly. Currently, few drugs such as Remdesivir (Med1), Convalescent plasma (Med2), Tocilizumab (Med3) and Hydroxychloroquine (Med4) are being used in a specified subgroup of patients. Remdesivir (Med1) can be prescribed for the infected persons with the moderate symptoms (those on oxygen) with limited contraindications. Convalescent plasma therapy (Med2) may be conducted to treat the patients with moderate disease who don’t respond (oxygen requirement is gradually increasing) despite the use of steroids. Tocilizumab (Med3) may be applied to patients with moderate disease and with constant requirements of oxygen and patients on mechanical ventilation who don’t show signs of improvement despite the use of steroids. Long-term safety therapies related to the treatment procedure of COVID-19 are still unknown to the greater extent. Hydroxychloroquine (Med4) has interacted in vitro activity against SARS-CoV2, and in several regional studies, it was said to be clinically beneficial although there were significant limitations of it. Those therapies are selected based on the effect of symptoms as well as their antiviral activity and possible side effects. The experts considered four major factors: antiviral activity (Sypt1), coolify (Sypt2), ease breathing (Sypt3) and side effect (Sypt4) as criteria for the performance evaluation of the therapies. The experts apply those therapies based on their past experience without strong evidence. They observed the functionalities of the therapies and realized the effect with certain positive and certain negative impacts and some unknown parts. The observed opinions (performance evaluation factors, antiviral activity (Sypt1), coolify (Sypt2), ease breathing (Sypt3) and side effect (Sypt4)) of the experts corresponding to the drugs (Remdesivir (Med1), Convalescent plasma (Med2), Tocilizumab (Med3) and Hydroxychloroquine (Med4)) for a particular patient are presented in this paper in the form of a picture fuzzy set. Those medicines are used for the treatment of COVID-19 affected patients based on the investigation. However, the Food and Drug Administration (FDA) has not approved yet those medicines as regular medicines. The FDA primarily gave permission to use those medicines for emergency service. The specified medicines can reduce some symptoms of COVID-19 affected patients who are simultaneously facing some other symptoms as a side effect. The medicine Tocilizumab (Med3) works effectively and the patient feels better sooner; at the same time, it increases the symptoms like cough or sore throat, block or runny nose, headaches or dizziness. To capture those kinds of uncertainties, we use PFN as it can manage uncertain situations using the neutral membership function. Henceforth, in this study, the decision matrix D is presented in the form of picture fuzzy information according to expert observation, which is shown in Table 9. This study assumes that the evaluation factors (Sypt1, Sypt2, Sypt3, Sypt4) corresponding to the four drugs (Med1, Med2, Med3, Med4) are represented using picture fuzzy numbers as in Table 9.

Table 9

Decision matrix D

	Antiviral activity (Sypt1)	Coolify (Sypt2)	Ease breathing (Sypt3)	Side effect (Sypt4)
Med1	[0.21, 0.48, 0.3]	[0.36, 0.23, 0.35]	[0.33, 0.35, 0.32]	[0.32, 0.34, 0.3]
Med2	[0.25, 0.4, 0.25]	[0.22, 0.35, 0.33]	[0.26, 0.23, 0.45]	[0.39, 0.26, 0.32]
Med3	[0.23, 0.33, 0.31]	[0.61, 0.22, 0.17]	[0.4, 0.1, 0.3]	[0.4, 0.3, 0.3]
Med4	[0.58, 0.13, 0.28]	[0.1, 0.2, 0.6]	[0.1, 0.41, 0.45]	[0. 2, 0.3, 0.2]

According to the definition of FUSH operation in Sect. 3.2, FUSH operation is performed among the PFNs of the four criteria corresponding to each alternative and the collective decision matrix is generated which is displayed in Table 10.

Table 10

Collective decision matrix

Alternative	Collective grade
Med1	[0.17, 0.02, 0.17]
Med2	[0.11, 0.01, 0.17]
Med3	[0.18, 0.02, 0.10]
Med4	[0.05, 0.00, 0.14]

Next the comparability sequence is computed from the collective decision matrix by the grey relational generating process which is defined in (6), (7) and (8). Generated compatibility sequence is presented in Table 11.

Table 11

Comparability sequence of the collective decision matrix D

Alternative	Compatibility sequence
Med1	[0.92, 0.0, 0.0]
Med2	[0.46, 0.5, 0.0]
Med3	[1.0, 0.0, 1.0]
Med4	[0.0, 1.0, 0.43]

Then we calculate the grey relational coefficient based on the membership grade of the therapies with respect to the symptoms. This is done using the expression (9) based on the compatibility sequence. The resultant grey relational coefficient is shown in Table 12. During the execution, we have considered the distinguished coefficient value as 0.5.

Table 12

Grey relational coefficient

Alternative	Grey relational coefficient
Med1	[0.08, 1.0, 1.0]
Med2	[0.54, 0.5, 1.0]
Med3	[0.0, 1.0, 0.0]
Med4	[1.0, 0.0, 0.57]

Next, the grey relation grade of the various membership degrees (positive, neutral and negative) is computed. Finally, the actual grey relation grades of the therapies are obtained by adding the grey relation grades of the respective membership degrees using Eq. (10). Here, the weights of all symptoms are considered to be equal. The resultant actual grey relation grades are displayed in Table 13.

Table 13

Grey relational grade of the alternatives

Alternative	Grey relational grade	Actual grey grade
Med1	[0.93, 0.5, 0.5]	1.93
Med2	[0.65, 0.67, 0.5]	1.82
Med3	[1.0, 0.5, 1.0]	2.5
Med4	[0.5, 1.0, 0.64]	2.14

Finally, the therapies are ranked according to the grey relational grade. The more be the grade, more will be the rank of the therapy or drug.

Therefore, the ranking order of the therapies is as follows, Med3 > Med4 > Med1 > Med2, i.e. Tocilizumab (Med3) will be more applicable for that particular patient in the process of treatment.

Comparison

The grey relation grades/scores and the final sequence of the therapies according to the different methods are shown in Table 14. According to the comparison table information, one can easily realize that the proposed method generates more accurate, clear and non-conflicting results, whereas the previously developed aggregation operator and cross-entropy-based MCDM methods may create the conflict situation. In the aggregation operator-based MCDM method, the score value is generated by the relation , which does not differentiate two PFNs and when and , because at this condition . Similarly, the entropy distance of two alternatives based on PFNs given in Wei (2016) will be equal, i.e. when and . Our proposed evidence-based MCDM approach manages this type of situation easily. The grade of the therapies depends on the evidence of the supporting grades. Suppose and , are the two PFNs where indicate different evidence and make different relational grades but generate same score value. We have compared the proposed method with three existing methods and shown that the generated ranking sequence of the proposed decision-making method is better than the ranking sequence of other methods.

Table 14

Comparative analysis with other methods

Decision approaches	Actual grey relational grade (Sypt1, Sypt2, Sypt3, Sypt4)	Ranking sequence
PFWA operator (Garg 2017)	(0.32, 0.32, 0.06, 0.29)	(Med2 > Med1 > Med4 > Med3 or Med1 > Med2 > Med4 > Med3)
Cross-entropy (Wei 2016)	(0.82, 0.82, 0.67, 0.47)	(Med4 > Med3 > Med1 > Med2 or Med4 > Med3 > Med2 > Med1)
TODIM method (Wei 2018a)	(0.72, 0.73, 0.87, 0.86)	(Med3 > Med4 > Med2 > Med1)
Our approach	(1.93, 1.82, 2.50, 2.14)	(Med3 > Med4 > Med1 > Med2)

According to the calculated actual grey relational grade of the medicines based on our proposed method, med is found to be better with respect to the remaining medicines. As per our study, the respective actual grey relational grades of the four applicable medicines are 1.93, 1.82, 2.50 and 2.14. Based on the actual grey relational grade, we can rank the medicines and choose the preferable one for better treatment. Moreover, we apply three other MADM methods, such as PFWA operator-based study (Garg 2017), cross-entropy method (Wei 2016) and TODIM method (Wei 2018a) on the same dataset and the corresponding results are shown in Table 14. The results shown in Table 14 depict that PFWA operator (Garg 2017) and cross-entropy method (Wei 2016) generate conflict situation, whereas TODIM method (Wei 2018a) generates a fair ranking among the medicines with narrow margin. But our proposed method estimates actual grey relational grade of each medicine and ranks them with strong evidence without any conflict incidence.

Validity testing

Wang and Triantaphyllou (2008) considered three generalized criteria to measure the acceptability of the newly proposed MCDM method. Newly developed methods may generate high-quality output, but the standard of the proposed method is measured by satisfying the following three criteria.

Criteria 1 The final ranking of the effective MCDM method does not change the best or optimal alternative due to the interchange of a non-optimal alternative by the worse alternatives without modifying the relative importance of each decision criteria.

Criteria 2 Transitivity property should be maintained by the effective MCDM method.

Criteria 3 The MCDM problem is divided into two subproblems and applies the same MCDM method to solve each of the subproblems and generate the rank of the alternatives as to the solutions. The resultant rank of the alternatives after combining the rank of the subproblems should be the same as the ranking of the original problem.

To check the validity of the proposed evidence-based MCDM method for medicine selection in case of COVID-19 infection, we have verified the given three criteria for validity testing by exchanging the parameters of optimal and worst alternatives, dividing the main problem into two subproblems and solving them, and finally checking the transitivity property by comparing the solutions of the subproblems.

Validity check of the proposed approach of criteria 1

This study has checked the validity of the proposed approach by the criteria 1, where the decision matrix given in Table 9 has been modified by the interchanging of the positive membership and neutral membership degree of the therapies Med1 (non-optimal alternative) with Med2 (worse alternative) for all the symptoms (Sypt1, Sypt2, Sypt3, Sypt4) and obtained the intermediate decision matrix D, which is shown in Table 15. The relative importance among the criteria (symptoms) is as usual. Then we have applied the proposed evidence-based MCDM method and generated the actual grey relation grades of the therapies (Med1, Med2, Med3, Med4), which are 2.17, 1.92, 2.50, and 2.00, respectively. Hence the ranking of the therapies according to the actual grey relation grade is (Med3, Med4, Med1, Med2), where the best therapy is Med3 which is also the best alternative in the original decision-making problem. Thereafter, it is proved that the proposed method does not change the ranking sequence of the therapies due to the interchange of positive membership and neutral membership degrees of those two therapies, where one is non-optimal and another is a worse alternative. Hence, the proposed method fulfils the criteria1.

Table 15

Modified decision matrix (D)

	Sypt1	Sypt2	Sypt3	Sypt4
Med1	[0.48, 0.21, 0.3]	[0.23, 0.36, 0.35]	[0.35, 0.33, 0.32]	[0.34, 0.32, 0.3]
Med2	[0.4, 0.25, 0.25]	[0.35, 0.22, 0.33]	[0.23, 0.26, 0.45]	[0.26, 0.39, 0.32]
Med3	[0.23, 0.33, 0.31]	[0.61, 0.22, 0.17]	[0.4, 0.1, 0.3]	[0.4, 0.3, 0.3]
Med4	[0.58, 0.13, 0.28]	[0.1, 0.2, 0.6]	[0.1, 0.41, 0.45]	[0. 2, 0.3, 0.2]

Validity check of the proposed approach of criteria 2 and criteria 3

To check the validity of the criteria 2 and criteria 3, the provided MCDM problem is decomposed into two subproblems with therapies (Med1, Med3, Med4) and therapies (Med1, Med2, Med3). Then, the two subproblems are solved by the proposed evidence-based MCDM method and the respective resultant ranks (Med3 > Med4 > Med1) and (Med3 > Med1 > Med2) are generated which are shown in Table 16. Then, the resultant sequence of the two subproblems is merged and the final sequence is obtained as (Med3 > Med4 > Med1 > Med2) which is equal to the ranking sequence of the original problem which satisfies the transitivity property. Hence the proposed method is valid based on criteria 2 and criteria 3 according to the established concepts given in Wang and Triantaphyllou (2008).

Table 16

Observation result of two subproblem

Subproblem	Actual grey relational grade	Ranking sequence
(Med1, Med3, Med4)	(Med1, 5.28), (Med3, 10.83), (Med4, 7.13)	(Med3 > Med4 > Med1)
(Med1, Med2, Med3)	(Med1, 4.58), (Med2, 3.10), (Med3, 10.83)	(Med3 > Med1 > Med2)

Conclusion

In this paper, we have developed an evidence-based medicine selection procedure that belongs to the probabilistic-based uncertainty for the treatment of COVID-19 patients. In the process, we have used PFS to represent the uncertain information, D–S theory to measure the probabilistic uncertainty of the neutral membership grade of PFS and GRA to measure the performance among the set of parameters that are in conflict and contradiction with each other. FUSH operation has been proposed for aggregating the PFNs. The evidence of neutral membership grade of the PFNs might be associated with positive or negative membership grades since the evidence of positive and negative membership grades is completely known. During the fusion, the evidence of the criteria is measured by the D–S theory. Then, we have calculated the evidence and non-evidence degree of the alternatives according to the resultant information using the basic probability assignment function. The contradictions among the criterion are managed by GRA using three normalization process as maximizing, averaging and minimizing the positive, neutral and negative membership grades, respectively. Next, the actual grey relation grade for each therapy is estimated using the membership wise grey correlation coefficient. Finally, the decision has been taken according to the actual grey relational grade of the therapies (alternatives). The proposed method is applied to find the preferable medicine for the treatment of COVID-19 patients. Due to the high mutation power of the COVID-19 virus, no approved drugs are found so far for better treatment. This method has successfully evaluated the preferences of the medicines based on the symptoms and signs of the COVID-19 patients. We have compared the proposed method with the existing three methods and resolved the conflict situation present in aggregation as well as entropy-based MCDM methods. This study has also checked the standard of the proposed method by satisfying the three generalized criteria successfully to measure the acceptance. In future, researchers can extend this model to other extensions of fuzzy sets such as rough sets and utilize the interdependency among the various evaluation criteria for better judgement. As the hidden information in neutral membership grades can be well expressed using rough set theory, one can use it to determine the interdependency of neutral membership grade with the positive and negative membership grades.