Skip to contents Skip to Global Navigation Menu
  • KDCA
  • Contact us
  • E-Submission

PHRP : Osong Public Health and Research Perspectives

OPEN ACCESS. pISSN: 2210-9099. eISSN: 2233-6052
Search
Advanced Search

Articles

Page Path

Original Article

Deep learning-based prognosis of major adverse cardiac events in patients with acute myocardial infarction: a retrospective observational study in the Republic of Korea

Vungsovanreach Kong1orcid, Kyung Ah Kim2,3orcid, Ho Sun Shon4orcid
Osong Public Health and Research Perspectives 2025;16(4):333-347.
Published online: July 23, 2025

1Department of Big Data, Chungbuk National University, Cheongju, Republic of Korea

2Biomedical Research Institute, Chungbuk National University Hospital, Cheongju, Republic of Korea

3Department of Biomedical Engineering, College of Medicine, Chungbuk National University, Cheongju, Republic of Korea

4Medical Research Institute, School of Medicine, Chungbuk National University, Cheongju, Republic of Korea

Corresponding author: Ho Sun Shon Medical Research Institute, School of Medicine, Chungbuk National University, 1 Chungdae-ro, Seowon-gu, Cheongju 28644, Republic of Korea E-mail: shon0621@gmail.com
Co-Corresponding author: Kyung Ah Kim Biomedical Research Institute, Chungbuk National University Hospital, and Department of Biomedical Engineering, College of Medicine, Chungbuk National University, 1 Chungdae-ro, Seowon-gu, Cheongju 28644, Republic of Korea E-mail: kimka@chungbuk.ac.kr
• Received: April 14, 2025   • Revised: June 10, 2025   • Accepted: June 15, 2025

© 2025 Korea Disease Control and Prevention Agency.

This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

  • 1,237 Views
  • 68 Download
  • 1 Crossref
prev next

Full Article

Download PDF Download PDF
  • Objectives
    This study developed deep neural network (DNN) models capable of accurately classifying major adverse cardiac events (MACE) in patients with acute myocardial infarction (AMI) after hospital discharge, across 3 follow-up intervals: 1, 6, and 12 months.
  • Methods
    DNN models were constructed to predict post-discharge MACE across 4 categories. Multiple traditional machine learning models were implemented as controls to benchmark the performance of our DNN approach. All models were evaluated based on their ability to predict MACE occurrence during the specified follow-up periods.
  • Results
    The DNN models demonstrated superior predictive performance over conventional machine learning methods, achieving high accuracies of 0.922, 0.884, and 0.913 for the 1-month, 6-month, and 12-month follow-up periods, respectively.
  • Conclusion
    The high accuracy of our DNN models highlights their practical advantages for AMI diagnosis and guidance of follow-up treatment. These models can serve as valuable decision support tools, enabling clinicians to optimize the overall management of AMI patients and potentially enhance their hospitalization experience.
  • Keywords: Cardiovascular; Deep learning; Early diagnosis; Major adverse cardiac events
The National Health Insurance Service of Korea (NHIS) analyzed trends in cardiovascular disease treatment over the past 5 years (2018–2022), reporting a 19.9% increase in patient numbers and a 38.5% rise in medical expenses related to heart disease during this period. The proportion of patients with cardiovascular disease increased across all age groups, with a particularly notable rise among individuals aged 30 years and below [1].
Acute myocardial infarction (AMI) occurs when the coronary arteries are blocked by a blood clot, cutting off blood supply to the heart muscle [2]. Approximately one-third of patients die before reaching the hospital immediately after onset. Rapid reopening of the occluded coronary artery is therefore critical to AMI management [3]. If a patient can be transported to a hospital capable of coronary intervention within 2 to 3 hours of symptom onset, balloon dilatation and stent insertion are recommended as standard treatments. Once stabilized and discharged, patients should receive secondary preventive treatments to reduce the risk of recurrence and complications. Secondary prevention includes optimal pharmacologic management for coronary artery disease. For patients with comorbid diabetes, hypertension, or hyperlipidemia, comprehensive self-management—such as smoking cessation, dietary changes, regular exercise, and stress management—should be undertaken alongside medication. However, even after specialized acute care and discharge, AMI patients often face challenges in resuming appropriate physical activity due to tachycardia, frequent arrhythmias, reduced cardiac output, orthostatic hypotension, and exercise-induced hypertension [4].
Numerous studies have applied deep learning algorithms to the study of cardiovascular diseases. For example, fundus photographs of the retina and ocular vasculature have been used in end-to-end deep learning models to directly predict cardiovascular diseases such as stroke and myocardial infarction (MI). In addition, a deep survival learning model was developed to account for the patient follow-up period and cardiovascular disease occurrence using fundus photography data [5]. Another study implemented an artificial intelligence (AI) model for detecting AMI from asynchronous electrocardiogram (ECG) data, comparing its performance with that of commercial ECG analysis software [6]. Lee et al. [7] further demonstrated the utility of deep learning models for predicting cardiovascular mortality and admission among patients with hypertension using the Korean National Health Information Database, showing improved performance relative to traditional statistical approaches. Rim et al. [8] advanced non-invasive risk stratification by developing a deep-learning system to predict coronary artery calcium scores from retinal photographs. In acute care settings, Liu et al. [9] used machine learning (ML) for variable selection and prediction of major adverse cardiac events (MACE) in emergency department patients presenting with chest pain. The rapid progress and potential of advanced analytical techniques in cardiovascular medicine have been summarized in comprehensive reviews, such as Krittanawong et al. [10], which offer practical primers on deep learning applications. Our study builds on this expanding literature by specifically applying deep neural networks (DNNs) to the multiclass prediction of MACE in post-AMI patients, using comprehensive clinical registry data. Recent work has also explored cardiovascular disease prediction via deep learning with feature augmentation [11]. The present study developed deep learning-based models to classify MACE in AMI patients for 3 follow-up periods: 1, 6, and 12 months. The Korea Acute Myocardial Infarction Registry (KAMIR) registry provides a robust data foundation for cardiovascular research, and DNNs are increasingly recognized for their strong predictive capabilities. The unique contribution of our study lies in the specific, multifaceted prognostic challenge addressed: using the KAMIR dataset to develop DNN models for nuanced, multiclass prediction of 4 MACE categories. An innovative feature of this research is the evaluation of predictive performance across 3 distinct and clinically relevant follow-up intervals. Additional novelty stems from our comprehensive comparison of the tailored DNN architecture with 6 traditional ML algorithms, each rigorously tuned for this complex classification task. Importantly, the application of Shapley additive explanations (SHAP) analysis to ensure interpretability of the DNN model’s MACE predictions in post-AMI patients marks a significant step toward practical clinical utility, distinguishing our work from more generalized approaches. The remainder of this paper is organized as follows: Section 2 presents the materials and methods, Section 3 details the results of both ML and deep learning models, and Section 4 discusses the findings in depth.
This section outlines the methodology for developing a DNN to classify MACE in patients. First, we describe the processes of data collection, filtration, and preprocessing. We then detail the feature selection approach, followed by a description of our DNN model architecture. Finally, we explain how the DNN’s performance was assessed and evaluated using various metrics. The aim was to classify MACE in AMI patients at 3 specific follow-up intervals: 1, 6, and 12 months, and to categorize MACE at each interval into 1 of 4 categories: coronary artery bypass grafting (CABG), cardiac death (CD), MI, and repeat percutaneous coronary intervention (Re-PCI). Distinct models were trained for each period to optimize classification accuracy. Figure 1 presents the experimental workflow, illustrating the processing of AMI patient data, including preprocessing, feature selection, model training, prediction, and evaluation.
Study Population
This study utilized data from the KAMIR. The dataset was constructed from AMI patients admitted to the hospital, incorporating a variety of patient information collected during hospitalization—including admission assessments, prior medical history, clinical measurements, cardiac interventions, and post-procedure follow-up data. Initially, we obtained 25,195 patient samples, each with 303 features. To ensure data quality, we applied predefined criteria to filter out irrelevant or outlier samples. First, patients who were pronounced dead on arrival were excluded. Next, we removed samples with excessive missing information, defined as lacking more than 70% of features, to avoid negative impacts on model development [12]. Finally, we excluded any cases lacking complete follow-up data for MACE—the primary target variable—at 1, 6, or 12 months. The extraction and filtering process is summarized in Figure 2, which displays the number of excluded samples at each step and the remaining dataset size. Through this multi-step process, 10,091 samples were excluded, resulting in a final cohort of 15,104 patients for analysis. The primary inclusion criterion was patients diagnosed with AMI who were admitted to the hospital. To ensure data quality and suitability for analysis, a multi-step filtration process was implemented (Figure 2). First, 278 patients dead on arrival were removed, followed by 3,438 cases with excessive missing data, and 6,375 lacking complete MACE follow-up for the target periods. This systematic filtering yielded a robust experimental dataset of 15,104 samples for further analysis, including feature selection, preprocessing, and model training.
Baseline clinical and demographic characteristics of the final patient cohort are detailed in Table 1. The table compares characteristics across the entire study population and by post-discharge MACE status, including groups with no MACE and those who experienced CABG, CD, MI, or Re-PCI. For example, patients who experienced CD were on average older (75.72±10.20 years) than those without MACE (65.77±12.68 years) and had a lower mean left ventricular ejection fraction (43.48%±13.46% vs. 53.53%±22.27%). Such differences underscore the clinical heterogeneity of the cohort and highlight variables potentially associated with MACE outcomes.
Data Preprocessing
Comprehensive data preprocessing was performed to enhance dataset quality and ultimately improve model performance. An essential first step was data cleaning, which involved verifying and correcting temporal and structural errors in the dataset [13]. Missing values were imputed using the mean for numerical variables and the mode for categorical variables. Duplicate entries were identified and removed to ensure data uniqueness. Exploratory data analysis (EDA) was also carried out to detect outliers and better understand the overall data structure [14]. Outliers identified during EDA were removed prior to model training. For rigorous and unbiased evaluation, the dataset was randomly divided into training (70%), validation (15%), and testing (15%) subsets. The training set was used for model fitting, enabling the DNN to learn patterns in the data. The validation set supported hyperparameter tuning and overfitting detection, while the independent testing set provided an unbiased assessment of final model performance on unseen data.
Feature Selection
It is common for medical datasets to include a large number of features, such as clinical variables and gene expression data. However, too many features can cause overfitting and unnecessary computational complexity in deep learning models. Therefore, feature selection techniques are essential for improving classification performance by removing duplicate, irrelevant, or noisy features. To identify the most informative variables, we used 5 feature selection methods: analysis of variance (ANOVA), t-distributed stochastic neighbor embedding (t-SNE), least absolute shrinkage and selection operator (LASSO), recursive feature elimination (RFE), and principal component analysis (PCA). ANOVA is effective for identifying variables that differ significantly between groups, making the process more targeted and efficient [15]. The LASSO technique removes less relevant features by shrinking their coefficients to zero [16]. RFE works by sequentially eliminating features that have the least impact on model performance [17]. In contrast, t-SNE reduces dimensionality by grouping similar data points together in a lower-dimensional space [18], while PCA transforms the data into a new set of principal components, also reducing dimensionality [19]. These feature-selection techniques, which were chosen to optimize the variance, are linear combinations of the initial variables. To systematically compare these techniques, we experimented with subsets of 20, 30, and 40 features, evaluating the resulting impact on model accuracy and performance. This process enabled a direct assessment of the efficacy of each feature selection strategy. After extensive testing, RFE with 30 features was found to yield the highest predictive accuracy on a dedicated validation set. This combination was thus selected for all subsequent ML and deep learning models presented in this study. Such comprehensive experimentation was crucial for understanding how feature selection methods and the number of features affect model predictive ability.
Synthetic Minority Over-sampling Technique
Synthetic minority over-sampling technique (SMOTE) addresses imbalanced datasets in ML by generating synthetic samples for minority classes, thereby improving class distribution without discarding valuable information [20]. SMOTE operates by selecting samples from the minority class and identifying their nearest neighbors in feature space; new synthetic samples are then created along the lines connecting these points, diversifying the representation of the minority class. This approach helps to reduce model bias toward the majority class and enhances predictive performance for minority classes [21]. By integrating SMOTE into data preprocessing—potentially in combination with undersampling the majority class—a more balanced dataset is achieved. This comprehensive strategy enhances both the generalizability and accuracy of classification models when facing imbalanced data [22]. As demonstrated in Table 2, the KAMIR dataset exhibits significant imbalance across all MACE classes.
Traditional Machine Learning
A variety of traditional ML algorithms were used as benchmarks to compare with the proposed DNN model. These algorithms included multinomial logistic regression (LR) [23], random forest (RF) [24], decision trees [25], support vector machines (SVM) [26], k-nearest neighbors (KNN) [27], and extreme gradient boosting (XGBoost) [28]. Each of these ML approaches was selected for its unique method of data analysis, ranging from linear models like LR, which estimates associations between variables, to ensemble methods such as RF and XGBoost, which construct multiple decision trees to improve prediction accuracy. SVM classifies data by finding the optimal hyperplane that separates different classes, while KNN assigns class labels based on the closest training samples in the feature space. To ensure the models’ generalizability to independent data, cross-validation was employed, thereby mitigating overfitting and supporting the selection of models with the best predictive capability [29]. All ML algorithms were fine-tuned using GridSearch, and Table 3 summarizes the optimal hyperparameters for each algorithm.
Proposed Deep Learning Architecture
Our DNN model was designed for multiclass classification tasks, specifically tailored for tabular clinical datasets. The architecture combines dense layers, batch normalization, rectified linear unit (ReLU) activation, and dropout regularization, allowing the model to efficiently process and learn from high-dimensional data [30]. The network begins with an input layer followed by 3 sequential dense layers, which are essential for capturing the underlying patterns within the dataset. Batch normalization is applied after each dense layer to stabilize and speed up the training process by normalizing outputs [31]. ReLU activation introduces nonlinearity, enabling the model to learn more complex relationships without excessive computational cost [32]. To reduce the risk of overfitting, dropout layers are placed after each ReLU activation, randomly omitting a portion of neuron outputs during training [33]. This encourages the model to develop generalized features that perform well on new, unseen data. Consequently, the model could learn more generalized features that are robust across new, unseen data. The output layer consists of 4 neurons with a SoftMax activation function, facilitating classification into 1 of 4 possible categories based on the highest predicted probability. The Adam optimizer is employed for training due to its adaptive learning rate and established performance in deep learning applications, while categorical cross-entropy serves as the loss function to maximize classification accuracy [34]. The DNN’s architecture thus achieves a balance between complexity and efficiency, ensuring strong performance on multiclass classification tasks using tabular data.
Specifically, the architecture starts with an input layer and proceeds through 3 dense layers with 512, 256, and 128 neurons, respectively. This narrowing structure was empirically chosen to capture hierarchical feature representations while controlling for overfitting. Batch normalization and ReLU activation are applied after each dense layer, while a dropout rate of 0.25, selected through preliminary tuning, provides robust regularization for this dataset and model configuration. By integrating these strategies, the model is able to efficiently learn and generalize, as depicted in the block diagram shown in Figure 3.
Model Evaluation
To comprehensively assess classification performance, we used a set of evaluation metrics appropriate for multiclass problems, including accuracy, precision, recall, and area under the receiver operating characteristic curve (AUC) [35]. For greater insight into the decision-making processes of the models, we incorporated explainable AI (XAI) techniques, with a focus on SHAP values [36]. This approach ensures that model predictions are transparent, reliable, and clinically meaningful, providing actionable insights for medical professionals. We conducted a detailed comparison of all models, evaluating predictive performance, interpretability, and computational efficiency to determine the most effective approach for MACE classification. The standard formulas used for each evaluation metric are as follows, where TP denotes true positives, TN true negatives, FP false positives, and FN false negatives:
(1)
Accuracy=TP+TNTP+TN+FP+FNPrecision=TPTP+FPRecall=TPTP+FNF1=2×Precision×RecallPrecision+Recall
This section presents the results of our experiments using both traditional ML algorithms and the proposed DNN models for classifying patient MACE. All models were evaluated using the metrics described previously. The section is organized as follows: Section 3.1 presents the outcomes of the traditional ML models, while Section 3.2 details the results of the proposed DNN models.
Machine Learning Performance Evaluation
To assess the effectiveness of the proposed DNN models, we conducted a comparative analysis using traditional ML algorithms as baseline controls. Each baseline model was trained with optimal hyperparameters, as summarized in Table 3. The primary objective was to evaluate the ability of each model to accurately classify MACE in patients discharged after AMI. The results of this analysis, which compare the performance of both the DNN and traditional ML models on this classification task, are systematically reported in Tables 4–6. These tables provide a comprehensive overview of model performance, allowing for a clear comparison of their effectiveness in predicting MACE and managing post-discharge outcomes for AMI patients.
As shown in Tables 4–6, the SVM model achieved the highest scores for metrics such as accuracy, recall, and F1-score, whereas the RF model outperformed others in terms of precision and AUC values. The use of SMOTE notably improved overall model performance by effectively addressing data imbalance through over-sampling. However, the remaining ML algorithms did not demonstrate satisfactory efficacy in classifying MACE in this patient population.
Proposed DNN Performance Evaluation
The proposed DNN models exhibited excellent performance in classifying MACE among AMI patients. Performance metrics for each follow-up period are provided in Tables 7–9, with the highest accuracy observed in the 1-month post-discharge classification. These tables present outcomes across various sets of training hyperparameters, highlighting the substantial benefits of hyperparameter optimization. Fine-tuning was found to be critical for identifying the best parameter configuration and maximizing model reliability for MACE classification.
Figure 4 illustrates the loss and accuracy curves for the DNN models on both training and validation datasets. These plots detail the changes in loss and accuracy over training epochs, illustrating that the model underwent substantial improvements during the first 15 epochs, followed by gradual stabilization as training continued. This trend suggests an initial phase of rapid learning, with performance plateauing as the model converged toward optimal parameters. Figure 5 presents the confusion matrices of the DNN models, showcasing the top-performing model selected for each follow-up period after hospital discharge.
Research Contributions
The primary motivation for this study was to address the critical need for accurate and timely prognostication of diverse MACE in patients following AMI across distinct post-discharge periods. Accurate prediction at 1, 6, and 12 months can significantly aid clinicians in tailoring follow-up intensity, optimizing therapeutic interventions, and improving overall patient outcomes. While previous research has explored ML in cardiovascular disease, this study specifically aimed to develop and validate DNN models capable of distinguishing among 4 MACE categories at these 3 key follow-up intervals. A further distinction of our work is the comprehensive benchmarking of DNN models against a range of traditional ML algorithms on a large-scale registry dataset (KAMIR), coupled with an XAI SHAP approach to provide interpretability of model predictions. This strategy addresses both predictive power and interpretability, offering clinicians valuable insight into the reasoning behind each prediction.
Performance Comparison
We observed a substantial disparity in performance between the ML models and the proposed DNN models. The DNN consistently outperformed all ML algorithms across every evaluation metric discussed in the previous section. Figures 6–8 visually represent the performance of all models in this experiment, specifically for MACE classification at 1, 6, and 12 months. The marked gap in performance underscores the advanced predictive capabilities of deep learning over traditional ML approaches for long-term outcomes in AMI patients.
Analysis of Feature Importance
Further insights were obtained from the deep learning models through the application of the XAI technique known as SHAP. Figure 9 illustrates the relative importance of each feature in predicting MACE after discharge in AMI patients. The SHAP summary plots display how each feature contributes to the model’s output, which is the probability of MACE at 1, 6, and 12 months. Features are ranked by the average magnitude of their SHAP values, indicating their impact on the model’s predictions. For classifying 1-month MACE, the most important features were age, left ventricular ejection fraction, and coronary angiography (CAG) post-thrombolysis in myocardial infarction (TIMI) score. At 6 months, the top features were left ventricular ejection fraction, age, and CAG post-TIMI score. For 12-month MACE, age, Killip class, CAG pre-TIMI score, and left ventricular ejection fraction emerged as the most significant predictors. These findings suggest that these features play a crucial role in determining the probability of MACE in AMI patients. Importantly, while these are the most influential features, all variables included in the model may contribute to the final prediction.
This study proposed a DNN model specifically designed to classify MACE in patients with AMI following discharge. The DNN classified patient MACE into 1 of 4 categories: CABG, CD, MI, and RE-PCI. Separate DNN models were developed to classify MACE at 3 follow-up intervals: 1, 6, and 12 months. The DNN models demonstrated superior classification performance compared to traditional ML approaches, representing a meaningful advancement in diagnostic and follow-up strategies for AMI patients. Cardiac clinicians may leverage these models to proactively assess MACE risk, thereby optimizing patient management and hospitalization. Despite these advances, the proposed DNN model has several limitations. Currently, it processes only clinical data and does not incorporate image data, which may further enhance MACE prediction accuracy. Additionally, the model was trained exclusively on the KAMIR dataset, predominantly sourced from South Korean hospitals. Its generalizability could be improved by including datasets from a wider range of international institutions. In future work, we plan to expand our research to include image-based deep learning models to further improve MACE classification accuracy. We also aim to develop an end-user application that enables healthcare providers without programming expertise to utilize these models in clinical practice. We believe our research represents a significant step toward improving post-discharge monitoring and care for AMI patients, thereby contributing to advancements in cardiac healthcare.
This study developed deep neural network models to predict major adverse cardiac events in patients following discharge for acute myocardial infarction (AMI). The models achieved high accuracy rates of 92.2%, 88.4%, and 91.3% for 1-, 6-, and 12-month follow-up periods, respectively, outperforming traditional machine learning approaches. These results validate the practical utility of deep neural networks for AMI diagnosis and post-discharge treatment planning. The models provide medical professionals with valuable tools to optimize patient management and improve the overall hospitalization experience.

Ethics Approval

Not applicable.

Conflicts of Interest

The authors have no conflicts of interest to declare.

Funding

This research was supported by the Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education (No. 2020R1I1A1A01065199 and RS-2023-00245300).

Availability of Data

The datasets are not publicly available but are available from the corresponding author upon reasonable request.

Authors’ Contributions

Conceptualization: all authors; Data curation: all authors; Formal analysis: VK; Funding acquisition: KAK, HSS; Investigation: KAK, HSS; Methodology: VK, HSS; Project administration: KAK, HSS; Resources: KAK, HSS; Software: VK; Supervision: KAK, HSS; Validation: KAK, HSS; Visualization: VK, HSS; Writing–original draft: VK, HSS; Writing–review & editing: all authors. All authors read and approved the final manuscript.

Figure 1.
Development of DNN models for classifying MACE in AMI patients.
DNN, deep neural network; MACE, major adverse cardiovascular events; AMI, acute myocardial infarction; EDA, exploratory data analysis; ANOVA, analysis of variance; RFE, recursive feature elimination; PCA, principal component analysis; t-SNE, t-distributed stochastic neighbor embedding; LASSO, least absolute shrinkage and selection operator; SMOTE, synthetic minority over-sampling technique; XAI, explainable artificial intelligence; SHAP, Shapley additive explanations; DT, decision tree; KNN, k-nearest neighbors; LR, logistic regression; RF, random forest; SVM, support vector machine; XGB, extreme gradient boosting; DNN, deep neural network.
Figure 1. Development of DNN models for classifying MACE in AMI patients.  
Figure 2.
Sample filtration for reducing non-relevant and outlier samples from the dataset. KAMIR, Korea Acute Myocardial Infarction Registry.
Figure 2. Sample filtration for reducing non-relevant and outlier samples from the dataset. KAMIR, Korea Acute Myocardial Infarction Registry.  
Figure 3.
Architecture of proposed deep neural network models using fully connected layers. ReLU, rectified linear unit.
Figure 3. Architecture of proposed deep neural network models using fully connected layers. ReLU, rectified linear unit.  
Figure 4.
(A–F) Loss and accuracy history of the best deep neural network (DNN) model for training and validation dataset for each discharge period. MACE, major adverse cardiovascular events.
Figure 4. (A–F) Loss and accuracy history of the best deep neural network (DNN) model for training and validation dataset for each discharge period. MACE, major adverse cardiovascular events.  
Figure 5.
(A–C) Confusion matrices of well-performing DNN models across post-hospitalization periods.
DNN, deep neural network; CABG, coronary artery bypass grafting; CD, cardiac death; MI, myocardial infarction; Re-PCI, repeat percutaneous coronary intervention.
Figure 5. (A–C) Confusion matrices of well-performing DNN models across post-hospitalization periods.  
Figure 6.
Performance comparison of machine learning and deep learning models (1-month MACE). MACE, major adverse cardiovascular events.
Figure 6. Performance comparison of machine learning and deep learning models (1-month MACE). MACE, major adverse cardiovascular events.  
Figure 7.
Performance comparison of machine learning and deep learning models (6-month MACE).
MACE, major adverse cardiovascular events; AUC, area under the receiver operating characteristic curve; KNN, k-nearest neighbors; DT, decision tree; LR, logistic regression; RF, random forest; SVM, support vector machine; XGB, extreme gradient boosting; DNN, deep neural network.
Figure 7. Performance comparison of machine learning and deep learning models (6-month MACE).  
Figure 8.
Performance comparison of machine learning and deep learning models (12-month MACE).
MACE, major adverse cardiovascular events; AUC, area under the receiver operating characteristic curve; KNN, k-nearest neighbors; LR, logistic regression; DT, decision tree; RF, random forest; XGB, extreme gradient boosting; SVM, support vector machine; DNN, deep neural network.
Figure 8. Performance comparison of machine learning and deep learning models (12-month MACE).  
Figure 9.
(A–C) Feature importance analysis using SHAP for all discharge periods.
SHAP, Shapley additive explanations; CAG, coronary angiography; TIMI, thrombolysis in myocardial infarction; LVEF, left ventricular ejection fraction; CD, cardiac death; CABG, coronary artery bypass grafting; Re-PCI, repeat percutaneous coronary intervention; MI, myocardial infarction; MACE, major adverse cardiovascular events.
Figure 9. (A–C) Feature importance analysis using SHAP for all discharge periods.  
Deep learning-based prognosis of major adverse cardiac events in patients with acute myocardial infarction: a retrospective observational study in the Republic of Korea
Table 1.
Baseline characteristics of KAMIR dataset used for MACE classification
Table 1.
Characteristic Post-discharge MACE
Total No MACE CABG CD MI Re-PCI
Total 15,104 (100.0) 14,317 (94.8) 44 (0.3) 263 (1.7) 171 (1.1) 546 (3.6)
Age (y) 65.91±12.71 65.77±12.68 66.93±10.03 75.72±10.20 68.83±13.93 65.75±11.95
BMI (kg/m2) 24.03±3.22 24.04±3.21 23.90±2.74 22.79±3.35 23.40±3.39 24.26±3.13
Systolic blood pressure (mmHg) 130.09±27.39 130.07±27.29 125.39±25.72 123.35±30.80 129.87±28.64 132.04±27.08
Diastolic blood pressure (mmHg) 79.16±16.11 79.16±16.04 77.64±15.21 75.26±18.35 78.72±15.67 80.41±16.30
PCI stent size (mm) 23.50±7.73 23.47±7.76 22.88±10.86 24.51±6.48 23.39±7.59 24.82±7.20
PCI stent diameter (mm) 3.17±0.47 3.17±0.46 3.27±0.54 3.12±0.58 3.13±0.41 3.11±0.45
CK (U/L) 1,133.97±1,777.96 1,142.49±1,784.77 819.12±877.10 1,167.22±2,130.81 718.67±1,223.75 1,196.75±1,876.80
CK-MB (U/L) 112.83±217.43 111.26±210.17 109.02±148.23 107.91±212.51 91.63±144.59 111.14±213.07
LVEF 53.41±21.94 53.53±22.27 44.10±11.94 43.48±13.46 49.77±13.38 53.72±11.65
CAG pre-TIMI 1.17±1.27 1.17±1.27 1.03±1.28 1.22±1.27 1.07±1.18 1.17±1.24
CAG post-TIMI 2.89±0.45 2.90±0.44 2.43±1.08 2.74±0.72 2.78±0.67 2.92±0.36
HDL cholesterol (mg/dL) 43.65±16.27 43.71±16.18 41.27±8.94 42.96±16.67 44.20±29.78 42.88±13.90
LDL cholesterol (mg/dL) 115.41±40.39 115.54±40.56 113.47±40.06 103.11±39.89 113.57±37.40 114.06±38.57
Sex
 Female 10,897 (72.1) 10,355 (72.3) 33 (75.0) 160 (60.8) 113 (66.1) 395 (72.3)
 Male 4,157 (27.5) 3,915 (27.3) 11 (25.0) 103 (39.2) 58 (33.9) 150 (27.5)
Chest pain
 No 11,413 (75.6) 3,303 (23.1) 9 (20.5) 113 (43.0) 47 (27.5) 160 (29.3)
 Yes 3,435 (22.7) 10,787 (75.3) 35 (79.5) 135 (51.3) 119 (69.6) 382 (70.0)
Killip class
 I 10,274 (68.0) 9,814 (68.5) 32 (72.7) 94 (35.7) 95 (55.6) 373 (68.3)
 II 1,810 (12.0) 1,691 (11.8) 4 (9.1) 32 (12.2) 29 (17.0) 67 (12.3)
 III 1,073 (7.1) 989 (6.9) 2 (4.5) 82 (31.2) 25 (14.6) 47 (8.6)
 IV 559 (3.7) 519 (3.6) 4 (9.1) 24 (9.1) 10 (5.8) 21 (3.8)
CAG target vessel
 LAD 6,398 (42.4) 6,038 (42.2) 10 (22.7) 90 (34.2) 79 (46.2) 244 (44.7)
 LCX 2,352 (15.6) 2,248 (15.7) 11 (25.0) 29 (11.0) 18 (10.5) 100 (18.3)
 Left main 262 (1.7) 241 (1.7) 5 (11.4) 15 (5.7) 4 (2.3) 6 (1.1)
 RCA 4,448 (29.4) 4,256 (29.7) 12 (27.3) 63 (24.0) 36 (21.1) 174 (31.9)

Data are presented as n (%) or mean±standard deviation.

KAMIR, Korea Acute Myocardial Infarction Registry; MACE, major adverse cardiovascular event; CABG, coronary artery bypass grafting; CD, cardiac death; MI, myocardial infarction; Re-PCI, repeat percutaneous coronary intervention; BMI, body mass index; PCI, percutaneous coronary intervention; CK, creatine kinase; CK-MB, creatine kinase MB; LVEF, left ventricular ejection fraction; CAG, coronary angiography; TIMI, thrombolysis in myocardial infarction; HDL, high-density lipoprotein; LDL, low-density lipoprotein; LAD, left anterior descending artery; LCX, left circumflex artery; RCA, right coronary artery.

Table 2.
Count of patient samples for each MACE type and each discharge duration
Table 2.
Period Patient sample (n)
Total CABG CD MI Re-PCI
1 mo 370 22 150 97 101
6 mo 468 15 143 46 264
12 mo 398 7 137 36 218
Total 1,236 44 430 179 583

MACE, major adverse cardiovascular event; CABG, coronary artery bypass grafting; CD, cardiac death; MI, myocardial infarction; Re-PCI, repeat percutaneous coronary intervention.

Table 3.
Best hyperparameter configuration for each ML algorithm after fine-tuning
Table 3.
ML algorithm Best hyperparameters
Decision tree max_depth=5; min_samples_leaf=10; criterion='gini'
Random forest n_estimators=100; max_depth=4; max_features='sqrt'
Support vector machines C=10; kernel='rbf'; gamma='scale'; degree=3
K-nearest neighbor n_neighbors=5; weights='uniform'; algorithm='auto'
Logistic regression C=0.1; penalty='l2'; solver='liblinear'
XGBoost n_estimators=50; learning_rate=0.1; max_depth=3; gamma=0

ML, machine learning; XGBoost, extreme gradient boosting.

Table 4.
Classification performance of traditional machine learning models (1 month)
Table 4.
ML model SMOTE ACC RE PR F1 AUC
SVM No 0.8254 0.8125 0.8395 0.8258 0.8712
Yes 0.8423 0.8615 0.8246 0.8425 0.8854
DT No 0.7915 0.8054 0.7789 0.792 0.8461
Yes 0.8136 0.8425 0.7856 0.8132 0.8654
RF No 0.8365 0.8215 0.8521 0.8366 0.8823
Yes 0.8554 0.8754 0.8356 0.8553 0.9015
LR No 0.8023 0.7954 0.8095 0.8024 0.8495
Yes 0.8256 0.8456 0.8054 0.8254 0.8685
KNN No 0.8512 0.8354 0.8675 0.8514 0.9005
Yes 0.8725 0.8954 0.8502 0.8723 0.9154
XGB No 0.7542 0.7754 0.7335 0.754 0.8154
Yes 0.7756 0.8054 0.7456 0.7754 0.8325

ML, machine learning; SMOTE, synthetic minority over-sampling technique; ACC, accuracy; RE, recall; PR, precision; F1, F1-score; AUC, area under the receiver operating characteristic curve; SVM, support vector machine; DT, decision tree; RF, random forest; LR, logistic regression; KNN, k-nearest neighbors; XGB, extreme gradient boosting.

Table 5.
Classification performance of traditional machine learning models (6 months)
Table 5.
ML model SMOTE ACC RE PR F1 AUC
SVM No 0.8235 0.8156 0.8421 0.8287 0.8876
Yes 0.8362 0.8475 0.8254 0.8363 0.8954
DT No 0.7952 0.7821 0.8095 0.7957 0.8523
Yes 0.8105 0.8234 0.7978 0.8105 0.8679
RF No 0.8176 0.8054 0.8302 0.8177 0.8754
Yes 0.8254 0.8395 0.8112 0.8252 0.8836
LR No 0.8025 0.7954 0.8102 0.8028 0.8615
Yes 0.8195 0.8321 0.8065 0.8193 0.8782
KNN No 0.8315 0.8254 0.8376 0.8314 0.8905
Yes 0.8392 0.8435 0.8354 0.8394 0.8985
XGB No 0.7854 0.7725 0.7985 0.7855 0.8425
Yes 0.7954 0.8095 0.7815 0.7954 0.8536

ML, machine learning; SMOTE, synthetic minority over-sampling technique; ACC, accuracy; RE, recall; PR, precision; F1, F1-score; AUC, area under the receiver operating characteristic curve; SVM, support vector machine; DT, decision tree; RF, random forest; LR, logistic regression; KNN, k-nearest neighbors; XGB, extreme gradient boosting.

Table 6.
Classification performance of traditional machine learning models (12 months)
Table 6.
ML model SMOTE ACC RE PR F1 AUC
SVM No 0.8254 0.8012 0.8423 0.8215 0.8652
Yes 0.8417 0.8495 0.8352 0.8423 0.8789
DT No 0.7925 0.7568 0.8211 0.788 0.8365
Yes 0.8103 0.8054 0.8154 0.8104 0.8492
RF No 0.8075 0.7895 0.8256 0.8073 0.8521
Yes 0.8231 0.8412 0.8054 0.823 0.8654
LR No 0.7854 0.7625 0.8095 0.7858 0.8312
Yes 0.8015 0.7954 0.8076 0.8015 0.8456
KNN No 0.8156 0.8025 0.8289 0.8157 0.8585
Yes 0.8324 0.8456 0.8195 0.8325 0.8712
XGB No 0.7542 0.7254 0.7854 0.7549 0.8054
Yes 0.7715 0.7654 0.7776 0.7715 0.8185

ML, machine learning; SMOTE, synthetic minority over-sampling technique; ACC, accuracy; RE, recall; PR, precision; F1, F1-score; AUC, area under the receiver operating characteristic curve; SVM, support vector machine; DT, decision tree; RF, random forest; LR, logistic regression; KNN, k-nearest neighbors; XGB, extreme gradient boosting.

Table 7.
Classification performance of DNNs for various hyperparameters (1 month)
Table 7.
Hyperparameter SMOTE ACC RE PR F1 AUC
Epoch=50, batch_size=64, learning_rate=1e-5 No 0.9054 0.9154 0.8956 0.9055 0.9425
Yes 0.9185 0.9354 0.9025 0.9184 0.9554
Epoch=50, batch_size=64, learning_rate=1e-4 No 0.9085 0.9254 0.8925 0.9084 0.9465
Yes 0.9215 0.9454 0.8975 0.9214 0.9584
Epoch=50, batch_size=64, learning_rate=1e-3 No 0.8955 0.8854 0.9055 0.8955 0.9385
Yes 0.9125 0.9354 0.8902 0.9124 0.9515

DNN, deep neural network; SMOTE, synthetic minority over-sampling technique; ACC, accuracy; RE, recall; PR, precision; F1, F1-score; AUC, area under the receiver operating characteristic curve.

Table 8.
Classification performance of DNNs for various hyperparameters (6 months)
Table 8.
Hyperparameter SMOTE ACC RE PR F1 AUC
Epoch=50, batch_size=64, learning_rate=1e-5 No 0.8675 0.8595 0.8754 0.8674 0.9254
Yes 0.8795 0.8854 0.8735 0.8795 0.9385
Epoch=50, batch_size=64, learning_rate=1e-4 No 0.8725 0.8678 0.8782 0.8727 0.9285
Yes 0.8842 0.8915 0.8765 0.884 0.9421
Epoch=50, batch_size=64, learning_rate=1e-3 No 0.8654 0.8585 0.8732 0.8656 0.9234
Yes 0.8812 0.8874 0.8751 0.8813 0.9392

DNN, deep neural network; SMOTE, synthetic minority over-sampling technique; ACC, accuracy; RE, recall; PR, precision; F1, F1-score; AUC, area under the receiver operating characteristic curve.

Table 9.
Classification performance of DNNs for various hyperparameters (12 months)
Table 9.
Hyperparameters SMOTE ACC RE PR F1 AUC
Epoch=50, batch_size=64, learning_rate=1e-5 No 0.8954 0.8854 0.9056 0.8955 0.9354
Yes 0.9125 0.9254 0.8995 0.9124 0.9485
Epoch=50, batch_size=64, learning_rate=1e-4 No 0.8954 0.8754 0.9156 0.8953 0.9325
Yes 0.9035 0.9154 0.8915 0.9032 0.9415
Epoch=50, batch_size=64, learning_rate=1e-3 No 0.8875 0.8654 0.9085 0.8865 0.9254
Yes 0.8962 0.9054 0.8875 0.8963 0.9345

DNN, deep neural network; SMOTE, synthetic minority over-sampling technique; ACC, accuracy; RE, recall; PR, precision; F1, F1-score; AUC, area under the receiver operating characteristic curve.

  • 1. Health Insurance Review and Assessment Service (HIRA) (KR). Current status of heart disease treatment [Internet]. HIRA; 2023 [cited 2025 Mar 19]. Available from: https://www.hira.or.kr/bbsDummy.do?pgmid=HIRAA020041000100&brdScnBltNo=4&brdBltNo=11037 Korean.
  • 2. Gong FF, Vaitenas I, Malaisrie SC, et al. Mechanical complications of acute myocardial infarction: a review. JAMA Cardiol 2021;6:341−9.
  • 3. Salari N, Morddarvanjoghi F, Abdolmaleki A, et al. The global prevalence of myocardial infarction: a systematic review and meta-analysis. BMC Cardiovasc Disord 2023;23:206.
  • 4. Baviera M, Genovese S, Colacioppo P, et al. Diabetes mellitus duration and mortality in patients hospitalized with acute myocardial infarction. Cardiovasc Diabetol 2022;21:223.
  • 5. Son J, Kim J, Kong ST, et al. Leveraging the generalization ability of deep convolutional neural networks for improving classifiers for color fundus photographs. Appl Sci 2021;11:591.
  • 6. Han C, Song Y, Lim HS, et al. Automated detection of acute myocardial infarction using asynchronous electrocardiogram signals: preview of implementing artificial intelligence with multichannel electrocardiographs obtained from smartwatches: retrospective study. J Med Internet Res 2021;23:e31129.
  • 7. Lee SJ, Lee SH, Choi HI, et al. Deep learning improves prediction of cardiovascular disease-related mortality and admission in patients with hypertension: analysis of the Korean National Health Information Database. J Clin Med 2022;11:6677.
  • 8. Rim TH, Lee CJ, Tham YC, et al. Deep-learning-based cardiovascular risk stratification using coronary artery calcium scores predicted from retinal photographs. Lancet Digit Health 2021;3:e306−16.
  • 9. Liu N, Koh ZX, Goh J, et al. Prediction of adverse cardiac events in emergency department patients with chest pain using machine learning for variable selection. BMC Med Inform Decis Mak 2014;14:75.
  • 10. Krittanawong C, Johnson KW, Rosenson RS, et al. Deep learning for cardiovascular medicine: a practical primer. Eur Heart J 2019;40:2058−73.
  • 11. Garcia-Ordas MT, Bayon-Gutierrez M, Benavides C, et al. Heart disease risk prediction using deep learning techniques with feature augmentation. Multimed Tools Appl 2023;82:31759−73.
  • 12. Palanivinayagam A, Damasevicius R. Effective handling of missing values in datasets for classification using machine learning methods. Information 2023;14:92.
  • 13. Stang J, Hartvigsen T, Reitan J. The effect of data quality on data mining: improving prediction accuracy by generic data cleansing. In: Proceedings of the 15th International Conference on Information Quality; 2010 Nov 12‒14; Little Rock, AR. ICIQ; 2010.
  • 14. Dhany HW, Sutarman S, Izhari F. Exploratory data analysis (EDA) methods for healthcare classification. J Intell Decis Support Syst 2023;6:209−15.
  • 15. Girden ER. ANOVA: repeated measures. Sage; 1992.
  • 16. Tibshirani R. Regression shrinkage and selection via the lasso. J R Stat Soc Series B Stat Methodol 1996;58:267−88.
  • 17. Guyon I, Weston J, Barnhill S, et al. Gene selection for cancer classification using support vector machines. Mach Learn 2002;46:389−422.
  • 18. Van der Maaten L, Hinton G. Visualizing data using t-SNE. J Mach Learn Res 2008;9:2579−605.
  • 19. Pearson K. LIII. On lines and planes of closest fit to systems of points in space. London Edinburgh Dublin Philos Mag J Sci 1901;2:559−72.
  • 20. Chawla NV, Bowyer KW, Hall LO, et al. SMOTE: synthetic minority over-sampling technique. J Artif Intell Res 202;16:321−57.
  • 21. Blagus R, Lusa L. SMOTE for high-dimensional class-imbalanced data. BMC Bioinformatics 2013;14:106.
  • 22. Ramezankhani A, Pournik O, Shahrabi J, et al. The impact of oversampling with SMOTE on the performance of 3 classifiers in prediction of type 2 diabetes. Med Decis Making 2016;36:137−44.
  • 23. Bohning D. Multinomial logistic regression algorithm. Ann Inst Stat Math 1992;44:197−200.
  • 24. Ho TK. Random decision forests. In: Proceedings of 3rd International Conference on Document Analysis and Recognition; 1995 Aug 14‒16; Montreal, CA. IEEE; 1995. p. 278‒82.
  • 25. Charbuty B, Abdulazeez A. Classification based on decision tree algorithm for machine learning. J Appl Sci Technol Trends 2021;2:20−8.
  • 26. Cortes C, Vapnik V. Support-vector networks. Mach Learn 1995;20:273−97.
  • 27. Altman NS. An introduction to kernel and nearest-neighbor. Am Stat 1992;46:175−85.
  • 28. Chen T, Guestrin C. Xgboost: a scalable tree boosting system. In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining; 2016 Aug 13‒17; New York, NY. ACM; 2016. p. 785‒94.
  • 29. Kohavi R. A study of cross-validation and bootstrap for accuracy estimation and model selection. In: Proceedings of the International Joint Conference on Artificial Intelligence; 1995 Aug 20‒25; Montreal, CA. IJCAI; 1995. p. 1137‒45.
  • 30. Schmidhuber J. Deep learning in neural networks: an overview. Neural Netw 2015;61:85−117.
  • 31. Ioffe S, Szegedy C. Batch normalization: accelerating deep network training by reducing internal covariate shift. In: Proceedings of the International Conference on Machine Learning; 2015 Jul 6‒11; Lille, FR, PMLR. 2015. p. 448−56.
  • 32. Agarap AF. Deep learning using rectified linear units (ReLU). arXiv [Preprint] 2018 Mar 22. https://doi.org/10.48550/arXiv.1803.08375.
  • 33. Srivastava N, Hinton G, Krizhevsky A, et al. Dropout: a simple way to prevent neural networks from overfitting. J Mach Learn Res 2014;15:1929−58.
  • 34. Kingma DP, Ba J. Adam: a method for stochastic optimization. arXiv [Preprint] 2014 Dec 22. https://doi.org/10.48550/arXiv.1412.6980.
  • 35. Han J, Pei J, Tong H. Data mining: concepts and techniques. 3rd ed. Morgan Kaufmann; 2022.
  • 36. Lundberg SM, Nair B, Vavilala MS, et al. Explainable machine-learning predictions for the prevention of hypoxaemia during surgery. Nat Biomed Eng 2018;2:749−60.

Download Citation

Download a citation file in RIS format that can be imported by all major citation management software, including EndNote, ProCite, RefWorks, and Reference Manager.

Format:

Include:

Deep learning-based prognosis of major adverse cardiac events in patients with acute myocardial infarction: a retrospective observational study in the Republic of Korea
Osong Public Health Res Perspect. 2025;16(4):333-347.   Published online July 23, 2025
DOI: https://doi.org/10.24171/j.phrp.2025.0120
Download Citation

Download a citation file in RIS format that can be imported by all major citation management software, including EndNote, ProCite, RefWorks, and Reference Manager.
Format:
Include:
Deep learning-based prognosis of major adverse cardiac events in patients with acute myocardial infarction: a retrospective observational study in the Republic of Korea
Osong Public Health Res Perspect. 2025;16(4):333-347.   Published online July 23, 2025
DOI: https://doi.org/10.24171/j.phrp.2025.0120
Close

Figure

Related articles

PHRP : Osong Public Health and Research Perspectives

About

Browse articles 

Editorial policies

For authors  

phrp 사이트 QRCODE(URL : https://ophrp.org)
© Korea Disease Control and Prevention Agency.