Fuzzy cognitive map based approach for determining the risk of ischemic stroke

Abstract

Stroke is the third major cause of mortality in the world. The diagnosis of stroke is a very complex issue considering controllable and uncontrollable factors. These factors include age, sex, blood pressure, diabetes, obesity, heart disease, smoking, and so on, having a considerable influence on the diagnosis of stroke. Hence, designing an intelligent system leading to immediate and effective treatment is essential. In this study, the soft computing method known as fuzzy cognitive mapping was proposed for diagnosis of the risk of ischemic stroke. Non‐linear Hebbian learning method was used for fuzzy cognitive maps training. In the proposed method, the risk rate for each person was determined based on the opinions of the neurologists. The accuracy of the proposed model was tested using 10‐fold cross‐validation, for 110 real cases, and the results were compared with those of support vector machine and K ‐nearest neighbours. The proposed system showed a superior performance with a total accuracy of (93.6 ± 4.5)%. The data used in this study is available by emailing the first author for academic and non‐commercial purposes.

1 Introduction

Health is considered as a cause of concern for every individual, particularly since the rate of harmful diseases is becoming an increasing human threat. There is a need for timely and accurate disease diagnosis which can provide protection against existing and future disease threats. The developments in soft computing methods have significantly contributed in the management, diagnosis, treatment, and taking care of human in health systems [1]. The diagnosis of diseases is a very complex and ambiguous task in medical science, as the theoretical data analysis methods have not been very effective in obtaining appropriate medical information, however, computer‐aided analysis has proven to be a powerful tool in diagnosing different types of diseases. There are numerous methods using artificial intelligence, and they have been used for the presentation of the data. The fuzzy cognitive mapping technique may be regarded as a lesser known soft computing method [2], however, it is an effective tool for representing causative knowledge and interferences, and it is able to describe the explicit conceptualisation of each system, obtained by a combination of the major characteristics of fuzzy logic and neural network. This method proposes a conceptual model, and it is appropriate for representing the concepts without an exact different concept using the fuzzy cognitive maps (FCM), so that each concept reflects a characteristic or defined state of the structure. Complex system behaviour is explained in the form of system [3, 4]. Easy construction and flexibility of analysis as well as the designing of the systems, compatibility, supporting uncertain knowledge, relatively simple modelling, time effectiveness, and decision making in higher level have caused the expansion of the FCM utilisation [5, 6].

It has various applications in the scientific and industrial fields, such as expansion of the decision making in support systems, political and military sciences, software engineering, information systems, business, modelling, pattern recognition, robotics, computer science, management, forecasting, expert systems, management, and so on [7–11]. In addition, FCM is considered as a modern and extremely efficient method for diagnosis of diseases. It can be employed in diagnosis, forecasting, classification, and decision making [4]. Using FCM and non‐linear Hebbian learning (NHL) algorithm, various diseases have been diagnosed so far such as Autism [12], Parkinson [13], breast cancer [5, 14], and celiac disease [15].

In a previous study [12], the onset of childhood autism was predicted with regard to 23 major factors of the disease, such as enjoy being swung, take an interest in other children, climbing on things, and so on. NHL algorithm was used for increasing efficiency of FCM, as well as achieving the right answer and making the right decision. Finally, the disease was classified into three categories (definite autism, no autism, probable autism), and the system accuracy was reported by 79.9%. In another study [13], Parkinson mathematical modelling was provided based on six main factors of the disease, such as tremor, rigidity, posture, and so on. Then, the disease was classified into six stages (healthy, stage 1, stage 2, stage 3, stage 4, stage 5, and stage 6), the obtained results were compared and simulated with and without the use of NHL algorithm. In a study [5], breast cancer classification was proposed using a two‐level FCM, so that at the first level based on several risk factors (age of patient, family history, BMI, number of children etc.), the degree of risk was classified into three groups (low, medium, and high) using NHL algorithm on 40 patients, and then the accuracy of NHL‐FCM system of classification was evaluated using two standard models of viz. Gail and Tyrer‐Cuzick. At the second level, also using 70 mammograms, 30 screening features, and the data‐driven NHL algorithm were classified into three classes (normal, benign, and malignant). Finally, the outputs of the two levels of FCM were integrated using expert rules. In this method, total accuracy was obtained by 94.3%. Also, in one study [14], the classification of breast lesions was proposed based on ten major histological features in three groups of UDH, ADH, and DCIS on 86 cases. In this study, the accuracy of UDH classification was determined to be 88 and 86% for ADH and DCIS, respectively, UDH was considered as benign lesion, and ADH and DCIS were considered as malignant lesions. In another research [15], celiac disease was graded as A, B₁, and B₂ among 89 cases considering seven major determinant factors of the disease using FCM and possibilistic fuzzy C‐means clustering algorithm. NHL algorithm was used to increase functionality and accuracy of FCM classification. The accuracy obtained from this method was equal to 88.8 and 91% for A and B₁ and B₂, respectively.

Stroke is a life‐threatening disease which has been ranked as the third leading factor causing morbidity in the world. It is considered to be a common cause of neurological disorders and as a significant health problem globally. Since, stroke is an emergency medical condition, hence reducing the diagnosis time along with enhancing the accuracy of prediction and diagnosis plays a vital role in preventing irrevocable damage to the patient's health [16, 17]. Several artificial intelligence‐based methods have been investigated to improve predictive efficiency for treatment and prognosis of stroke occurrence.‏ Anindito Fnu et al. [18] using expert system prevented stroke based on known controllable factors such as smoking, blood pressure, physical activity, BMI, and so on. Gupta and Mishra [19] proposed diagnosis of ischemic stroke using MRI images and artificial neural network (ANN). This study consisted of six stages. The first stage involved the collection of MRI images, the second stage was concerned with preprocessing of MRI images, the third stage was finding the midline to achieve symmetric images, the fourth stage was dividing the images into the two regions of right and left, the fifth stage was extraction of features using grey level co‐occurrence matrix (GLCM) and finally the last stage was using ANN for normal and infected brain classifications. Singh and Choudhary [20] predicted the stroke using artificial intelligence. In this method, decision tree algorithm was used for selecting features, principal component analysis (PCA) algorithm was used for reducing dimensions and ANN was used for model classification. This method was compared with different methods available for predicting stroke on the dataset of CHS. The results of this model showed the efficiency of this model with accuracy of 97.7% compared to other methods. Islam et al. [21] presented a backup system for stroke prediction using fuzzy C‐means clustering algorithm for classifying data based on the risk factors, lost through a primary data mining process, and fuzzy inference system (FIS) and adaptive neuro fuzzy inference system (ANFIS) were used to generate fuzzy logic rules to achieve higher accuracy for early diagnosis of stroke. In another study [22], the stroke was predicted using nine predictive factors for stroke (hypertension, smoking, diabetes etc.). This prediction was done using the knowledge discovery process (KDP) method as a simple and low computational tool for discovering patterns along with ANN and support vector machine (SVM) models for extracting data patterns. These two extracted models were evaluated based on the findings of accuracy and area under curve (AUC). The obtained results showed that ANN has more predictive capability compared to SVM.

Possibly, stroke is not diagnosed in the first step so, medical diagnosis is important for prediction of future state of the patient's health. Given this, for the first time, in the present study the risk rate of ischemic strokes was investigated (after considering its prevalence and risk factors/features) based on FCM and NHL algorithm. The NHL algorithm was used to maximise the efficiency and flexibility as well as eliminating the disadvantages of FCM [13]. This algorithm is based on the Hebbian learning rule and is applied for improving the FCM structure [23].

In other words, a general solution was proposed, so that physicians can diagnose strokes in a timely manner. In this method, the stroke risk rate of the person was specified using real data samples, and the results were compared with the mean opinion of certain neurologists. The results were also compared with and without the use of NHL algorithm. Finally, the overall accuracy of the proposed method was computed and compared with the SVM and K ‐nearest neighbours (KNN) classifiers. It is believed that the presented method along with the simulation of several neurologists’ opinion has a better performance than that of a physician. This paper is structured as follows: Section 2 proposes a definition of fuzzy cognitive mapping. Section 3 describes NHL algorithm. Section 4 provides the information on the construction and implementation of the FCM model to determine the risk rate of ischemic stroke during the next 5 years, then the simulation results are compared and the system performance is evaluated, and finally, Section 5 concludes the research.

2 Fuzzy cognitive maps

The FCM is a soft computing method, trying to follow an inferential process similar to humans decision‐making and inference. It was proposed by Kosko for the first time in 1986, as indicative directional graphs with feedback loops, and was used for representing a computing dependency and complexity of the system. The FCM includes a set of concepts/factors, shown by $C_i$ where $i=1,\dots ,N$ (N is the total number of concepts). Each concept depicts a key factor of the system, and it is specified by the value $V_i$ belonging to the interval of [0, 1]. The concepts are connected to each other by weight arrows showing their relationship. A simple FCM with four factors and nine weight arrows is shown in Fig. 1. Any connection between the two concepts of $C_j$ and $C_i$ has a weight belonging to the interval of [−1, 1] and the value of this weight depends on the intensity of causative relation between $C_j$ and $C_i$. The relationship between these concepts is either positive or negative as depicted by $W_{ji}$. Accordingly, three weights can exist, $W_{ji}>0$ showing the positive relationship; $W_{ji}<0$ showing the negative relationship, and $W_{ji}=0$ showing no relationship between these concepts. Specifying the values of the concepts and weights, FCM converges to a stable condition. In each step, the value of each concept $V_i$ is influenced by the connected concepts and updated according to the following equation:

$$V_i^{\left(K+1\right)}=f\left(V_i^{\left(K\right)}+\sum _{j=1\cdot j\ne i}^N{V}_j^{\left(K\right)}{W}_{ji}\right)$$ (1)

where f is sigmoid threshold function, so that: $\phi \ge 0$, and it is the parameter determining the function slope and is used as follows:

$$f=\frac{1}{1+{\mathrm{e}}^{-\phi x}}$$ (2)

All values are calculated and the process is continued till the FCM is converged to a fixed state after some iteration [4, 11, 23].

Fig. 1.

Sample FCM with four factors

3 NHL algorithm

The NHL algorithm was proposed by E. Papageorgiou as a method for FCM model learning. The NHL is based on the Hebbian learning rule. The main principle of this algorithm depicts that all the concepts are triggered in FCM model in each iteration step and their values change concurrently. This algorithm updates the weights proposed by the experts. While applying the NHL algorithm, only the non‐zero weights proposed by the experts are updated in each of the iterations, the values of zero weights do not change, and new connections are not specified. The sign and direction of the calculated weights do not change. NHL is successfully used for training due to sustainability, rapid convergence, ease of use, being accurate, and as system response is close to the actual value. Combining the knowledge of experts with learning methods such as NHL makes FCMs proceeding to a high‐perspective control method [24, 25]. This algorithm uses the learning rate parameter (α) and the parameter of weight decay (β), in order to calculate the updated weights, thus the concepts’ values are changed. The Hebbian learning algorithm is shown according to the following equation:

$$W_{ji}^K=\text{sgn}\left(W_{ji}^{\left(K-1\right)}\right)\left(\beta \left|W_{ji}^{\left(K-1\right)}\right|+\alpha V_i^{\left(K-1\right)}(V_j^{\left(K-1\right)}-\text{sgn}\left(W_{ji}^{\left(K-1\right)}\right)W_{ji}^{\left(K-1\right)}{V}_i^{\left(K-1\right)}\right))$$ (3)

The values of each FCM concept and the value of each weight $W_{ji}^K$ are updated using (1) and (3). There are two different termination conditions determining the stop time of the algorithm. The first condition is concerned with the minimisation of the following objective function:

$$F_1=\sqrt{∥{\text{DOC}}_j^{\left(K\right)}-T_j^2∥}$$ (4)

$T_j$ is the final value of decision output concept (DOC), and value of DOC belongs to the interval of $\left[T_j^{\max }\cdot T_j^{\min }\right]$. The second termination condition is based on the consecutive changes of two values of $\text{DO}{\mathrm{C}}_j$ considered according to the following equation:

$$F_2=\left|{\text{DOC}}_j^{\left(K+1\right)}-{\text{DOC}}_j^{\left(K\right)}\right|<€$$ (5)

where € is the tolerance level minimising the changes in the DOC values. When the conditions for termination of the algorithm are met, the final weight matrix, W _NHL, is obtained [13, 15]. Algorithm flowchart is shown in Fig. 2.

Fig. 2.

NHL algorithm flowchart

4 Definition of stroke and introducing risk factors for creating FCM model

Stroke is a significantly serious neurological disease affecting all ages throughout the world. It is the third leading cause of morbidity after coronary disease and cancer in the USA with an occurrence rate of 0.2–2 per thousand people. In Iran, the morbidity rate caused by the stroke was reported by 8% in 2003, and the rate of lost lives was reported by 4.4% [26]. Stroke is classified into two main groups: (i) haemorrhagic strokes and (ii) ischemic strokes. The latter is the most common one, comprising 85–90% of strokes. It occurs when the blood circulation to a part of the brain is interrupted or reduced significantly, thus affecting the supply of oxygen and nutrients to that part. During that time, which may occur for a few minutes to several hours, the brain cells lose their function and die [27, 28]. Hence, this state is considered as a medical emergency, and prompt treatment is critical to reduce the damage and consequent neurological disability, indicating the importance of every second for diagnosis and treatment.

The ischemic stroke risk factors include 12 factors, divided into controllable and uncontrollable factors playing an important role in the diagnosis of this disease [29, 30] as shown in Fig. 3. Neurologists consider these risk factors in the diagnosis of the stroke after performing a physical examination and observing the results of individual tests. The values of these concepts consist of two, three, or four fuzzy values with low, medium, high, and very high linguistic variables as summarised in Table 1. For example, in this paper, three linguistic variables were for HDL cholesterol concept: low<35, 36<medium<60, and high>61. These input concepts were collected from the patients referred to Amiralmomenin Hospital in Iran, and were specified by three neurologists, namely Dr. Mohammadzad, Dr. Hagigat, and Dr. Asgarpour. This novel dataset can be freely provided, for academic purposes, by emailing to the first author. The FCM model proposed in this paper was based on the opinions of these neurologists who decided on the input and output concepts. Fig. 4 shows the membership functions of the risk rate for a stroke.

Fig. 3.

Risk factors of ischemic stroke

Table 1.

FCM model factors for diagnosis of ischemic stroke

Concepts	Type of values
C1 : age	three fuzzy values (young, middle age, old), young<45, middle age 46–65, 66>old
C2 : blood pressure	four fuzzy values (low, medium, high, very high); low<130, medium 131–150, high 151–170, very high>171
C3 : LDL cholesterol	four fuzzy values (low, medium, high, very high); low<130, medium 131–160, high 161–190,very high>191
C4 : HDL cholesterol	three fuzzy values (low, medium, high); low<35, medium 36–60, 61>high
C5 : diabetes	three fuzzy values (low, medium, high); low<70, medium 71–125, 126>high
C6 : heart disease	two discrete values (present, absent)
C7 : family history	two fuzzy values (yes, no)
C8 : smoking	two fuzzy values (yes, no)
C9 : BMI	three fuzzy values (low, medium, high); low<19, medium 20–25, 26>high
C10 : exercise	two fuzzy values (yes, no)
C11 : sex	two fuzzy values (female, male)
C12 : stroke history	two fuzzy vales (yes, no)
C13 : risk of stroke	three fuzzy values (low, medium, high)

Fig. 4.

Membership functions of output concept (C13)

4.1 FCM implementation for diagnosis of stroke risk rate

After determining the input and output concepts by the neurologists, they were asked to express the effect of each concept on the other concepts as well as the relationship between the concepts by linguistic variables in the fuzzy rule form of ‘if‐then’. These rules are used by experts to determine the relationship between concepts by linguistic variables (e.g. low, medium, and high). Each linguistic variable can belong to any value within the interval of [0, 1] [15]. For example, according to the neurologist's opinions, the relationship between blood pressure (C₂) and output concept (C₁₃) is defined as follows:

• The first neurologist: if the blood pressure is medium, then the risk rate is medium.

• The second neurologist: if the blood pressure is medium, then the risk rate is high.

• The third neurologist: if the blood pressure is medium, then the risk rate is very high.

The summation of three linguistic variables of medium, high, and very high was done using SUM method. Then, using the defuzzification technique in the centre of gravity method, the value resulting from C₂ to C₁₃ was obtained as 0.54 as illustrated in Fig. 5. Thus, in a similar fashion, all the initial weights of the FCM were obtained as shown in Table 2. The relationship between concepts and calculated weights for the proposed FCM model (given in Table 2) is shown in Fig. 6 regarding the prediction of risk rate for ischemic stroke.

Fig. 5.

Aggregation of three linguistic variables using the SUM method

Table 2.

Initial weights values proposed by neurologists

	C1	C2	C3	C4	C5	C6	C7	C8	C9	C10	C11	C12	C13
C1	0	0.55	0	0	0.35	0.4	0	0	0	0	0	0.60	0.60
C2	0	0	0	0	0.46	0.45	0	0	0	0	0	0.52	0.54
C3	0	0	0	0	0	0.44	0	0	0	0	0	0.40	0.40
C4	0	0	0	0	0	−0.55	0	0	0	0	0	−0.54	−0.58
C5	0	0.40	0	0	0	0.45	0	0	0	0	0	0.50	0.50
C6	0	0	0	0	0	0	0	0	0	0	0	0.55	0.58
C7	0	0.45	0	0	0.35	0.40	0	0	0	0	0	0.30	0.30
C8	0	0.30	0	0	0.20	0.35	0	0	0	0	0	0.45	0.45
C9	0	0.25	0	0	0.45	0.35	0	0	0	0	0	0.25	0.25
C10	0	−0.20	0	0	−0.35	−0.35	0	0	−0.30	0	0	−0.30	−0.30
C11	0	0	0	0	0	0	0	0	0	0	0	0.10	0.10
C12	0	0	0	0	0	0	0	0	0	0	0	0	0.68
C13	0	0	0	0	0	0	0	0	0	0	0	0	0

Fig. 6.

Proposed FCM model with the assigned initial values of weights for determining the risk of ischemic stroke

5 Results of simulation and discussion

For this problem, 12 characteristics or concepts as inputs and C₁₃ as output concept of decision making were considered by the neurologists for depicting the risk rate of an ischemic stroke during the next 5 years. The decision‐ making output concept is defined as a variable from the fuzzy set as low, medium, and high, which is based on the opinions of the neurologists as: 0 ≤ low≤0.15, 0.16≤medium≤0.31, and 0.32≤high≤1.

In designing the FCM model for stroke, the concepts’ values are updated concurrently till a final state is reached, according to (1). The initial concepts’ values are the original concepts of FCM, required for its implementation, playing important roles in the occurrence of stroke. In order to better illustrate this subject, two examples of the test data (1 male and 1 female) were provided in the following.

First example : In this example, the data related to a male with stroke history is given in Table 3, and the neurologists predicted a medium probability of stroke recurrence.

Table 3.

The data related to a male with stroke history for determining the risk of ischemic stroke

age	74
blood pressure	180
LDL	140
HDL	35
blood sugar	120
heart disease	0
family history	0
smoking	1
BMI	26.8
physical exercise	0
sex	1
stroke history	1

At first, the values are normalised in the interval of [0, 1] according to the following equation in [5]:

$$x_{\text{normalised}}=\frac{x-x_{\min }}{x_{\max }-x_{\min }}$$ (6)

After normalisation of the above values, the initial input values are obtained as: C _Initial  = [0.68 0.5 0.28 0 0 1 0.4 0 1 1 0]. These initial values and weight matrices shown in Table 2 are updated according to (1), until the FCM reaches a balanced point. As seen in Table 4, after seven iteration steps, the concept values do not change, depicting that the balance point is reached. After seven iterations, the output concept value reaches to 0.67759. According to (7) in [2], the risk rate is obtained by 35% and is considered as high risk based on the neurologists’ opinion. Thus, the NHL algorithm is used for setting the FCM initial weights to approximate the system response to the real value. In Fig. 7, the sequence diagram of concept values shows that the convergence (balance point) is reached

$$\text{Risk}\left(x\right)=\left\{\begin{array}{rl}& 0,x\le 0\cdot 5\\ & \frac{x-0\cdot 5}{0\cdot 5}\times 100\%,0\cdot 5<x\le 1\end{array}\right.$$ (7)

Table 4.

Values of FCM concepts at seven iteration steps

C1	C2	C3	C4	C5	C6	C7
0.68	0.5	0.36	0.25	0.28	0	0
0.54577	0.59026	0.52428	0.51687	0.57747	0.58257	0.5
0.53677	0.60125	0.53533	0.53483	0.59417	0.61708	0.5337
0.53617	0.60206	0.53607	0.53604	0.59492	0.61892	0.53596
0.53613	0.6021	0.53612	0.53612	0.59493	0.61899	0.53611
0.53613	0.6021	0.53613	0.53613	0.59493	0.61899	0.53612
0.53613	0.6021	0.53613	0.53613	0.59493	0.61899	0.53613

C8	C9	C10	C11	C12	C13
1	0.4	0	1	1	0
0.56709	0.52697	0.5	0.56709	0.65987	0.641
0.5382	0.52542	0.5337	0.5382	0.64803	0.67554
0.53626	0.52464	0.53596	0.53626	0.64765	0.6775
0.53613	0.52454	0.53611	0.53613	0.64761	0.67759
0.53613	0.52453	0.53612	0.53613	0.64761	0.67759
0.53613	0.52453	0.53613	0.53613	0.64761	0.67759

Fig. 7.

Subsequent values of concepts till convergence without applying NHL algorithm

5.1 Using NHL algorithm in FCM for determining stroke risk rate

This algorithm is used for improving the efficiency of FCM and better modelling of the system behaviour, so as to obtain acceptable results by training the FCM. To do so, the NHL algorithm is applied for training the FCM model, which is as follows:

• Step1 : Updating the input concepts’ values, decision‐making output concept value, and initial weight using (1) and (3) to obtain a stable state.

• Step2 : Meeting the final conditions according to (4) and (5), which is equal to 0.1 and 0.001, respectively.

• Step3 : Firstly, the learning rate of 0.1 is considered, and then this value reduces exponentially in each of the iterations of K until 0.001 is obtained according to (8) in [5]

  $${\alpha }^{\left(K\right)}=\frac{{\alpha }^{\left(K-1\right)}}{\left(2K+1\right)}$$ (8)

• Step4 : Determining weight coefficient value β using the trial and error method [23]. After applying the training process in the first example, the stroke risk rate reaches a value of 0. 6273 after 25 iterations, depicting a risk rate of 25%, which belongs to the interval of 0.16≤medium≤0.31 according to the neurologists’ opinion, and is considered as medium risk. In Fig. 8, the sequence diagram of concept values shows the process till convergence by applying NHL algorithm. Second example : In this example, based on the female patient's data, the neurologists forecasted a low recurrence of stroke as given in Table 5.

  Fig. 8.

  

  Subsequent values of concepts till convergence with applying NHL algorithm

Table 5.

The data related to a female without stroke history for determining the risk of ischemic stroke

age	64
blood pressure	120
LDL	100
HDL	45
diabetes	85
heart disease	1
family history	1
smoking	1
BMI	26.8
physical exercise	0
sex	0
stroke history	0

After normalisation of the above values according to (6), the following initial concepts values are obtained: C _Initial  = [0.53 0.2 0.19 0.5 0.13 1 1 1 0.26 0 0 0 0]. After simulation by FCM, the initial values are presented as: C _FCM  = [0.52014 0.55601 0.52014 0.52014 0.55195 0.56495 0.52014 0.52014 0.51386 0.52014 0.52014 0.58043 0.59674]. After seven iterations, the output concept value converges to a value of 0.59674 depicting the risk rate of 19%. The NHL algorithm is applied so as to approach the system response to a real value and make an accurate decision. The weight matrix updated by NHL algorithm is shown in Table 6. Thus, after applying the NHL algorithm, the following final concepts values are obtained after 25 iterations: C _Final  = [0.5141 0.5370 0.5141 0.5141 0.5344 0.5428 0.5141 0.5141 0.5104 0.5141 0.5141 0.5522 0.5621]. The DOC value reaches a value of 0.5621, depicting that the risk of recurrent stroke is equal to12%, which is considered as low risk and located in the interval of 0 ≤ low≤0.15 according to the neurologist's opinion.

Table 6.

Updated weight matrix with NHL algorithm for the first example

	C1	C2	C3	C4	C5	C6	C7	C8	C9	C10	C11	C12	C13
C1	0	0.44	0	0	0.36	0.44	0	0	0	0	0	0.53	0.53
C2	0	0	0	0	0.40	0.44	0	0	0	0	0	0.44	0.48
C3	0	0	0	0	0	0.44	0	0	0	0	0	0.36	0.36
C4	0	0	0	0	0	−0.46	0	0	0	0	0	−0.47	−0.50
C5	0	0.36	0	0	0	0.35	0	0	0	0	0	0.44	0.44
C6	0	0	0	0	0	0	0	0	0	0	0	0.49	0.51
C7	0	0.40	0	0	0.31	0.37	0	0	0	0	0	0.27	0.27
C8	0	0.27	0	0	0.18	0.32	0	0	0	0	0	0.40	0.40
C9	0	0.23	0	0	0.40	0.31	0	0	0	0	0	0.22	0.22
C10	0	−0.16	0	0	−0.29	−0.29	0	0	−0.25	0	0	−0.25	−0.25
C11	0	0	0	0	0	0	0	0	0	0	0	0.09	0.09
C12	0	0	0	0	0	0	0	0	0	0	0	0	0.60
C13	0	0	0	0	0	0	0	0	0	0	0	0	0

The 10‐fold cross‐validation method was used to evaluate the proposed system using 110 real datasets in the age range of 28–95 years old. In this method, nine datasets were considered for training and one dataset was used for testing. In each implementation, the accuracy and recognition rate were calculated for the test dataset. After ten iterations, the total accuracy of the algorithm obtained from the mean of accuracies was equal to (93.6 ± 4.5)%. Table 7 depicts the accuracy and recognition rate in each algorithm iteration for the test dataset. The output value results of FCM for all datasets after applying NHL are shown in Fig. 9.

Table 7.

Proposed NHL‐FCM system evaluation results in ten iteration

	Accuracy	Recognition rate low	Recognition rate medium	Recognition rate high
fold 1	91%	1	0.8	1
fold 2	81%	0.66	1	0.75
fold 3	91%	1	0.8	1
fold 4	100%	1	1	1
fold 5	91%	0	0.75	1
fold 6	100%	1	1	1
fold 7	100%	1	1	1
fold 8	100%	1	1	1
fold 9	91%	1	0.85	1
fold 10	91%	0	1	1

Fig. 9.

FCM output values results for all people data after application of NHL

Also for making a better comparison, the data used in this paper were classified by SVM and KNN. SVM is widely used in medical and biological research to classify various diseases, which can map incoming vectors into a high‐dimensional space using several kernel functions [22]. Due to the good results of SVM in the various areas, it is known as a potential competitor for FCM. Free parameters of SVM, such as kernel length (C), were tuned by nested cross‐validation method. Table 8 shows the accuracy and recognition rate in each algorithm iteration for the SVM classifier. In addition to SVM, in order to further ensure about the accuracy of the proposed system, KNN classifier was also used. KNN classifier is one of the traditional classifiers used for machine learning, classifying a specific sample of the data of this classifier, so that K finds a sample closer to the data and acts on the basis of voting. So, each of those neighbours will regard those particular data as members of their class. Usually, Euclidean distance criterion is used to find the nearest sample K. The value of K is equal to 10 [31, 32]. Table 9 shows the accuracy and recognition rate in each algorithm iteration for the KNN classifier. The comparison of the values in Table 7 with those in Tables 8 and 9 showed that SVM in fold 2 with accuracy of 81% and low recognition = 0.66, medium recognition = 1, and high recognition = 0.75, and in fold 9 with accuracy of 91% and low recognition = 1, medium recognition = 0.85, and high recognition = 1, and KNN only in fold 5 with accuracy of 91% and low recognition = 0, medium recognition = 0.75, and high recognition = 1 are equal to FCM with better results. In other folds, FCM outshined SVM and KNN classifiers in terms of accuracy and recognition rate. As a result, it can be said that proposed NHL‐FCM model is stronger than SVM and KNN classifiers. Comparison of the values in Tables 8 and 9 showed that KNN only in fold 5 with accuracy of 91% and high recognition = 1 performs better than SVM, in fold 3 with accuracy of 81% and low recognition = 1, medium and high recognition = 0.8, and in fold 10 with accuracy of 81% and low recognition = 0, medium recognition = 1, and high recognition = 0.87 it performs similar to SVM, and in other 7 folds, SVM offers better results than KNN in terms of accuracy and recognition rate. Consequently, it can be said that the classification of these three methods with respect to the power of performance is as follows: FCM, SVM, and KNN.

Table 8.

System evaluation results with the SVM classifier in ten iterations

	Accuracy	Recognition rate low	Recognition rate medium	Recognition rate high
fold 1	81%	1	0.6	1
fold 2	81%	0.66	1	0.75
fold 3	81%	1	0.8	0.8
fold 4	91%	1	1	0.83
fold 5	81%	0	0.75	0.85
fold 6	91%	1	0.75	1
fold 7	91%	1	1	0.83
fold 8	91%	0.5	1	1
fold 9	91%	1	0.85	1
fold 10	81%	0	1	0.87

Table 9.

System evaluation results with the KNN classifier in ten iterations

	Accuracy	Recognition rate low	Recognition rate medium	Recognition rate high
fold 1	72%	1	0.6	0.8
fold 2	72%	0.66	1	0. 5
fold 3	81%	1	0.8	0.8
fold 4	81%	1	0.75	0.83
fold 5	91%	0	0.75	1
fold 6	81%	1	0.75	0.83
fold 7	81%	1	0.75	0.83
fold 8	81%	0.5	1	0.87
fold 9	81%	1	0.85	0.66
fold 10	81%	0	1	0.87

6 Conclusion

Early diagnosis and treatment of stroke is vital, since early diagnosis not only increases the chances of recovery and survival of the patient but also protects the patient against severe consequences of the stroke. In the present study, an efficient method was proposed involving a soft computing technique known as fuzzy cognitive mapping along with a NHL algorithm to predict the risk rate of ischemic stroke in next 5 years based on the main risk factors. The NHL algorithm was applied to enhance the FCM performance. Indeed, in this method, the rate of precise disease diagnosis increased by combining the experts’ knowledge and experiences with the fuzzy logic system. The purpose of this study is not to implement the system, but to use a predefined system and extend it to a new area in order to extract new knowledge in the field of stroke disease. Until now, no other paper has proposed the use of FCM for the classification of stroke patients. In other words, using this tool is for a superior purpose, which is the same as helping to extract knowledge in the field of diagnosing and predicting of stroke. To do so, the results obtained from the system performance were compared with the mean of the mentioned opinions of the neurologists. The overall accuracy obtained using the NHL‐FCM method for 110 actual dataset was compared with those of the SVM and KNN classifiers. The results indicated that the proposed model achieved an accuracy of (93.6 ± 4.5)%, which is higher than that of SVM and KNN. The results showed that modelling by FCM was very similar to the viewpoints of the neurologists. Therefore, it is believed that, the proposed method has the capability to evaluate the risk rate of stroke in next 5 years. Moreover, it can be said that the FCM is a flexible and reliable decision‐making method, which can be used by the neurologists as an excellent and efficient system for diagnosis of the stroke. This manuscript is the beginning of a long way of solving both of these problems.