This retrospective cohort study was conducted at the Department of Cardiovascular Medicine, the Second Affiliated Hospital of Nanchang University (a teaching tertiary hospital), in Jiangxi Province, China. We collected electronic medical records of patients with AMI who underwent PCI for the first time from January 2018 to December 2019. These patients were followed up through December 2021.

The inclusion criteria of the participants were as follows:

The following exclusion criteria were applied:

We adopted the “Guidelines for Developing and Reporting Machine Learning Predictive Models in Biomedical Research” to guide the reporting of our study [20].

Data were collected from electronic health records, including demographic characteristics, clinical investigations, the first laboratory tests, and disease-related events. MACEs were defined as cardiomyopathies (excluding infectious, familial, alcohol, and drug-related cardiomyopathies), hypertensive heart disease, recurrent myocardial infarction, heart failure, sudden cardiac death, revascularization, malignant arrhythmia, and stent thrombosis [21]. Abnormal Q waves were identified by the clinician based on ECG results. Left ventricular ejection fraction (LVEF) was defined as normal (more than 50%), mildly abnormal (40% to 50%), moderately abnormal (30% to 40%), and severely abnormal (less than 30%) [22]. According to the number of diseased coronary vessels and implanted stents, they were classified as I, II, III, and IV.

This study was reviewed and approved by the Second Affiliated Hospital of Nanchang University Medical Ethics Committee (No. Review 2017 No. (098)).

All analyses were performed with R software (version 4.0.1; R Core Team). The patients were randomly assigned to training (n=953, 70%) and testing (n=409, 30%) data sets by calling the createDataPartition function using the random number method, and chi-square tests showed that there was no statistical difference between them (χ2₁=2.169; P=.14). We developed machine learning models using the training data set. We analyzed the missing and out-of-range values with imputation methods. We used multiple imputation with chained equations to assign any missing predictor values [23]. The imputation processes were performed separately in the training and testing sets after the data were split. To improve the accuracy of the machine learning models and increase the speed of finding the optimal solution by gradient descent, we standardized and normalized all input variables before the model was built. To alleviate the problem of imbalanced classification samples, we adopted the random oversampling method. We used the ROSE package in R to generate new balanced training data. After random oversampling, the number of patients with MACE in the training data sets changed from 186 to 471.

The model was built using demographic information (age and sex), personal comorbidities (diabetes and peripheral arterial disease), preoperative PCI (LVEF, the number of diseased vessels, and abnormal Q waves), serological examination (beta 2 microglobulin, B-type brain natriuretic peptide, glucose, serum creatinine clearance, and estimated glomerular filtration rate), and the characteristics of PCI (the number of implanted stents; n=65; Table S1 in Multimedia Appendix 1). A total of 12 variables with significant differences in the univariate analysis were included in the model development (Table 1).

Based on a previous application of the model [24], the parameter range of the model was preset, and the GridSearchCV function was used to select the optimal parameters of each machine learning model.

To minimize potential overfitting in the above machine learning models, we called the trainControl function in the caret package of R language for 7-fold cross-validation during the development process. The model performance was assessed for accuracy, recall, precision, area under the receiver operating characteristic curve (AUC), and F₁-score in the testing data set. We identified the important predictors through importance analysis of the variables. Logistic regression analysis was used to compare the absolute value of the coefficients of variables; RF was used to measure the importance of features by calculating information gain through entropy; and the ANN method was used to calculate the relative importance of variables based on the generalized weight method.

The following R packages for machine learning approaches were used: caret, randomForest, and neuralnet. Baseline characteristics were compared with the Wilcoxon rank sum test for continuous variables and the chi-square test for categorical variables. We considered P<.05 (2-sided) to be statistically significant.

In this study, we developed a machine learning–based model integrating clinical, anatomical, and laboratory test features to predict MACEs in patients with newly diagnosed AMI after their first PCI. The major findings suggest that the ANN model had higher predictive accuracy (accuracy of 87.99%, AUC of 0.81, and F₁-score of 0.71), compared to RF, KNN, SVM, and logistic regression.

Among the patients with AMI in this study, the rates of MACEs at 30 days, 6 months, 1 year, and 3 years after PCI were 4.63%, 11.38%, 14.54%, and 18.50%, respectively, and the incidence of MACEs at 30 days after PCI was slightly less than the 5.5% reported in the Harmonizing Outcomes with RevascularIZatiON and Stents in Acute Myocardial Infarction study (HORIZONS-AMI) [25]. The incidence of MACEs at half a year was higher than the 6.67% reported by Chow et al [26], consistent with the 2-year rate of MACEs reported by Sanmenxia City (18.06%). The survival condition of patients with AMI after PCI was slightly different from that in other studies. The participants in this study were all patients who were first diagnosed with AMI and underwent PCI for the first time, and their prognosis was better than that of patients with previous myocardial infarctions and multiple PCIs [27]. In addition, the progression of a patient’s disease is affected by not only individual differences but also access to medical resources and services. The HORIZONS-AMI trial was first reported in 2012. Although the treatment level in the HORIZONS-AMI trial was higher than that available in China at that time, with the development of China’s economy, the progress of science and technology, and the substantial improvement of medical care, the MACE rate obtained in our study was lower than that reported in the HORIZONS-AMI study.

One study found that machine learning demonstrated the highest performance for risk prediction in patients with extracardiac vascular disease for the prediction of both arrhythmogenic cardiomyopathy and MACEs [10]. McCord et al [28] proposed that machine learning can be used to assess AMI within 30 minutes and that the algorithm has high diagnostic and prognostic utility. In this study, 3 algorithms were used to predict MACE occurrence for patients with newly diagnosed AMI undergoing PCI treatment for the first time. The MACE prediction ability of the logistic regression model was lower than that of the ANN model and almost the same as that of the RF model. However, the positive predictive values of these 3 prediction models were not high. Kuang et al [29] also found that the ANN model had the best predictive value for the transition from mild cognitive impairment to Alzheimer disease with ideal stability [29]. The positive predictive values of the RF model and the logistic regression model were both approximately 50%, which means that their predictive ability for MACEs was poor. Their shortcomings may be associated with class imbalances [30], which can easily cause the predicted results to be biased toward a large number of classes (the positive type of fault can be placed into the negative class). ANNs, with their powerful self-adaptability, self-organization, fault tolerance, and “black box” operation of nonlinear mapping, are especially suitable for solving problems with complex internal mechanisms and have been widely used in various disciplines [31].

Our results indicated that the 3-year prognostic risk among patients with AMI undergoing their first PCI was mainly related to age, ECG characteristics, ventricular ejection ability, coronary artery lesions, stent implantation after PCI, and some serological variables. Yang et al [32] found that the risk ratio of hospital deaths after PCI was 3.723 (95% CI 2.86-4.84) for South Korean patients aged >65 years relative to those aged ≤65 years. A Korean multicenter AMI National Institutes of Health–registered project found that the MACE rate, 3 years after PCI, among patients with AMI with an LVEF <40% was 3.34 times that of the control group [33]. Fam et al [34] conducted a retrospective study on patients with clinical AMI in Asian multiethnic groups and found that the risk of MACEs among patients with diabetes, 2 years after PCI, was 1.84 times higher than that among patients without diabetes [34]. Diabetes is a chronic metabolic disease, and long-term diabetes is often accompanied by bleeding disorders, vascular endothelial dysfunction, small artery lesions, high blood sugar [35], hemostatic disorders [36], endothelial dysfunction, and a series of other changes [37]. These characteristics will accelerate the process of atherosclerotic disease deterioration. The number of coronary artery lesions and the number of stents implanted in a patient are also positively correlated with the risk of postoperative MACEs to a certain extent. This may be because a higher number of vessels with coronary artery disease and the number of implanted stents tend to indicate a more serious condition, leading to a worse prognosis for the patients. Hongbo et al [38] found that the probability of a poor prognosis in patients with multiple coronary artery lesions was 20.0%, compared with 6.98% in patients with single coronary artery lesions [38].

The results of the machine learning model showed that predictors like LVEF, number of implanted stents, and age were more important to the model. LVEF is a common variable that reflects left ventricular function, and patients with a low LVEF have a significantly higher MACE rate [39]. An increase in age can lead to the aggravation of atherosclerosis [40]. The number of implanted stents may be related to the severity of the disease and the extent of the infarction [41]. This reminds us that we should pay special attention to the prognosis of patients with AMI who have a low LVEF value, older age, and more implanted stents in clinical practice.

This study has some limitations. First, there may have been an issue of survival bias in the study, as patients with missing follow-up data were excluded. Second, the data have missing values. We have filled missing values with multiple imputation; however, imputation with these techniques could synthetically reduce the variance in these variables and may have affected the accuracy of the constructed model. Finally, although the models were internally validated with data from the same hospital, further work should include validation with external data from other hospitals or centers.

This study revealed that the ANN model showed good MACE prediction performance for patients with AMI after PCI, and it identified the most important predictors, which may aid in clinical decision-making and improve outcomes. This model needs to be externally validated in larger populations and multicenter settings.