Prediction of Major Adverse Cardiovascular Events (MACCE) Following Percutaneous Coronary Intervention Using ANFIS-PSO Model

Background: This study aimed to use the hybrid method based on an adaptive neuro-fuzzy inference system (ANFIS) and particle swarm optimization (PSO) to predict the occurrence of major adverse cardiac and cerebrovascular events (MACCE) of patients underwent angioplasty. Method: This is a retrospective cohort study comprised a total of 220 patients (69 women and 151 men) who underwent coronary angioplasty in Ekbatan medical center in Hamadan city, Iran between March 2009 until March 2012. The occurrence and non-occurrence of MACCE, (including death, CABG, stroke, repeat revascularization) were considered as a binary outcome. The performance of ANFIS models for predicting MACCE was compared with ANFIS-PSO and logistic regression. Results: Ninety-six patients (43.6%) experienced the MACCE event after ten years of follow-up. In multivariate analysis based on logistic regression model, variables such as age (OR = 1.05), smoking (OR = 3.53), diabetes (OR = 2.17) and stent length (OR = 3.12) had a signicant effect on MACCE occurrence. Comparing the prediction performance of the models showed that the ANFIS-PSO model had higher accuracy (89%) compared to the ANFIS (81%) and logistic regression (72%) in the prediction of MACCE. Conclusion: The performance of ANFIS-PSO has a minimum error and maximum accuracy compared to other models in the prediction of MACCE. Application of this model is recommended for intelligent monitoring of these patients, the classication of high-risk patients and the allocation of necessary medical and health resources based on the needs of these patients. The ANFIS-PSO had a high accuracy in predicting the MACCE event compared to the ANFIS and the logistic regression models. Various studies have used classical and machine learning approaches to predict cardiovascular disease, for example: Taghizadeh et al., compared ANFIS and logistic regression in predicting death after coronary artery bypass graft surgery. Their study evaluated the data of 824 patients including age, sex, BMI, hypertension, diabetes (mellitus), blood cholesterol, peripheral vascular disease, addiction, smoking, history of chronic heart failure, etc as input variables. The FCM method was used to create FIS model, Gaussian and linear membership functions were used for inputs and output. Sensitivity, specicity, accuracy were reported ANFIS 67%, 97% and 96% and in logistic regression 48%, 89% and 89%, respectively


Introduction:
The most common cardiovascular disease is coronary artery disease, which presents as ischemia and acute myocardial infarction. It is estimated that more than 23.6 million deaths in various communities will be due to cardiovascular disease [1]. Coronary artery disease is the rst and most common cause of death in Iranians of all ages [2].
One of the major complications of the cardiovascular system is atherosclerosis or blockage of the arteries that supply blood to the heart. Bypass surgery, balloon angioplasty, and stent implantation are some of the most important therapeutic interventions used to treat clogged arteries [3]. A stent is a wired metal mesh tube that is used to keep the blood supply to the artery at the blocked point open. Angioplasty is a less invasive and cheaper procedure than coronary artery bypass graft surgery. In addition, it only requires 1 to 2 days of hospitalization, and the patient can return to work sooner and start his activity [4].
Non-fatal myocardial infarction events, cardiac death, and the need to restore blood ow are major adverse cardiovascular events that may occur to any patient after stent implantation. Due to the increasing demand for stent angioplasty, predicting the occurrence and identifying potential risk factors for subsequent major cardiovascular events can be effective in improving patients' survival and quality of life [5].
The development of an expert system, especially an adaptive neuro-fuzzy inference system (ANFIS) method, over the past few decades has made accurate predictions possible in many areas of medicine [6,7]. The most important advantages of these systems are: expression of human knowledge using special linguistic concepts and fuzzy rules, nonlinearity and compatibility, and better accuracy of these methods in terms of data constraints compared to other methods [8,9]. Another new modeling method is arti cial neural networks, the most important reason for the strength is their ability to be learned from input and output training patterns. The combination of fuzzy systems, which are based on logical rules, and arti cial neural networks with the ability to knowledge acquisition from numerical information, enables us to use human knowledge in the construction of the prediction model [10]. Therefore, this study aimed to predict the occurrence of major adverse cardiac (MACCE) in patients undergoing stent angioplasty using the ANFIS model.

Method:
This retrospective cohort study comprised 220 patients underwent coronary angioplasty in Ekbatan Medical Center in Hamadan city, Iran from March 2009 to March 2012. From the beginning of September 2020 after at least 10 years of follow-up, the clinical condition of patients was followed up by telephone interview or in-person contact or by using the patients' medical records from the treating physician's o ce or hospital.
Patients' information including demographic characteristics, number of involved vessels, the number and type of stent (metal, drug), stent size, height and length, history of smoking, history of diabetes, hypertension, and hyperlipidemia extracted from patient records. The occurrence and non-occurrence of MACCE (including death, CABG, stroke, repeat revascularization) were considered as a binary outcome.
The Risk score variable is de ned as the sum of risk factors including diabetes, smoking, hypertension, and hyperlipidemia. If the person has the desired symptom, the value is 1, and if the symptom not exists, the value takes 0, so when the value of the risk score is 2 for an individual, it means that the person has 2 risk factors.
Predictive models including ANFIS, PSO-ANFIS, and logistic regression were examined to predict the occurrence of MACCE. The odds ratio and 95% con dence interval were used to summarize and describe the data in the logistic regression model. ANFIS as a fuzzy Takagi-Sugno system is based on if-then rules and bene ts the learning skill of arti cial neural networks together with decision-making capability of fuzzy-logic [11]. The ANFIS structure is formed of ve layers. The rst layer performs the fuzzi cation process where the number and type of membership function are speci ed by a training system based on training data. Then, the rules are de ned in the rule layer and the effect of each rule is calculated, which can also be called the inference layer. In the third layer, the effect of each rule is normalized according to the effect of other rules. For the fourth layer, the output of each rule is obtained, which calculates the weighted output. Finally, in the fth layer, the outputs are added together to form the output of the fuzzy system. To create this network, several parameters such as the type of membership function, the number of functions, the learning method, and the number of repetitions (Epoch) must be optimized [11,12].
The Particle swarm optimization (PSO) is a stochastic optimization technique bio-inspired by the social behavior found in nature such as the motion of bird ocks and schooling sh. In this population-based algorithm, the members of the population interact straightly with each other and solve the problem by exchange of their best experience and recalling valuable experiences of the past. The beginning of the algorithm in optimizing the parameters of membership functions in the ANFIS model is that after designing the initial FIS, a group of particles or solutions are randomly generated and initialized by moving these particles during successive iterations and by updating the position of the particles, they try to nd the optimal solution to the problem. For each answer, the t value of each particle is evaluated, and if a better t value is obtained, the position of the particle is updated. In the next step, the best new position of the whole group is found, and if the stop criterion is met, the algorithm also stops [13,14].
The input variables used in predictive models including sex, age, stent length, stent diameter, number of vessels, type of stent and risk score. ANFIS architecture was shown in gure 1. For all models; the data were divided randomly into two groups of training (70%) and test (30%) sets. The same test, training sets and input variables were used to compare the performance of different models. Data analysis was performed using SPSS software version 21 and fuzzy logic toolbox of Matlab20 software. The stop criterion in the present study was a maximum of 1000 epochs or an error of less than 0.01. The best FIS was selected based on the lowest error rate.
Ethics approval, consent to participate This study was approved by the Ethics Committee of Hamadan University of Medical Sciences with IR.UMSHA.REC.1398.017. Due to the retrospective nature of the study, the patient consent form was not signed. All the participants were informed on the purpose of the study, in response to a follow-up phone call, patients or their relatives responded if they wished. All methods were performed in accordance with the relevant guidelines and regulations.

Results:
At the end of the follow-up time, out of 220 patients, 96 patients (43.6%) had experienced MACCE events. From those, 48 patients passed away (21.8%), 16 patients (7.3%) developed CABG, 5 patients had a nonfatal myocardial infarction (2.3%) and 27 patients (12.3%) required repeat revascularization. The results related to the distribution of different variables in terms of occurrence and non-occurrence of MACCE events are presented in Table 1.  Step Size Decrease Rate 0.9 Step Size Increase Rate 1.1 Error Goal 0 In the multivariate analysis based on the logistic regression model, the effect of age, smoking, diabetes, and stent length on the MACCE event was signi cant. In this regard, the results showed that a one-year increase at the age will increase the 5% chance of MACCE occurrence (95% CI: 1.02-1.09). Diabetes also increases the chance of developing MACCE by 5.77 times (95% CI: 2.05-16.202). The chance of developing MACCE in smokers is 3.5 times non-smokers (95% CI: 1.61-7.75). Also, the chance of MACCE occurrence in people who have a stent length greater than 20 mm is 3.12 times those in whom the stent length is less than 20 mm (95% CI: 1.48-6.57).
The performance of the different predictive models in the training and test phase is presented in a table.
The process of error changes during the training process for the ANFIS and ANFIS-PSO models is presented in Figure 2. Finally, a comparison of the prediction performance of three models in the form of sensitivity, speci city, and accuracy indices is presented in Table 4. Discussion: In this study, the long-term (10-year) survival status of cardiovascular patients undergoing angioplasty was evaluated. Also, to predict the occurrence of MACCE, the performance of logistic regression model as the most common classical model, ANFIS and ANFIS-PSO as a data mining model were compared.
One of the features of the present study is that it evaluates the long-term survival of patients, while other studies mainly consider short intervals (6 months). Also, in these studies, the main focus has been on comparing the survival of patients in two types of metal and drug stents. The present study was a cohort study in which no randomization was performed, so it was not possible to directly compare patient survival in two types of drug and metal stents. However, to modify its effect, this variable is added to the model. Farshidi et al., [20].
As expected, risk factors such as older age and smoking increase the risk of MACCE occurrence and are in accordance with the ndings of Farshidi et al., and Tsai et al., [19,20].
Also, patients with 3 stents implanted were 1.8 times more likely to have MACCE events than patients with 1 stent. This result was also con rmed in the Tsai et al., [19].
The ANFIS-PSO had a high accuracy in predicting the MACCE event compared to the ANFIS and the logistic regression models. Various studies have used classical and machine learning approaches to predict cardiovascular disease, for example: Taghizadeh et al., compared ANFIS and logistic regression in predicting death after coronary artery bypass graft surgery. Their study evaluated the data of 824 patients including age, sex, BMI, hypertension, diabetes (mellitus), blood cholesterol, peripheral vascular disease, addiction, smoking, history of chronic heart failure, etc as input variables. The FCM method was used to create FIS model, Gaussian and linear membership functions were used for inputs and output. Sensitivity, speci city, accuracy were reported ANFIS 67%, 97% and 96% and in logistic regression 48%, 89% and 89%, respectively [10].
In recent years, due to the emergence of hybrid prediction methods that help in screening and predicting the consequences of the disease, diagnosis and prediction in the eld of medicine has made signi cant progress. According to the performance of ANFIS-PSO more considerable than traditional ANFIS and logistic regression in the prediction of MACCE. Application of this model is recommended for intelligent monitoring of these patients, the classi cation of high-risk patients and the allocation of necessary medical and health resources based on the needs of these patients. Declarations:

Con ict Of Interest
There are no con icts of interest.

Funding
This study was supported by Vice-Chancellor of Research and Technology of Hamadan University of Medical Sciences, Contractor No. 980210772.

Figure 2
Legend not included with this version.