This paper is in the following e-collection/theme issue:

Artificial Intelligence (1369) Cardiology (157) Machine Learning (1463) Cardiac Risk and Cardiac Risk Calculators (45) Clinical Information and Decision Making (1404) Clinical Informatics (1018) Decision Support for Health Professionals (1184) New Methods (944) Advanced Data Analytics in eHealth (133)

Published on in Vol 26 (2024)

Preprints (earlier versions) of this paper are available at https://preprints.jmir.org/preprint/49848, first published .
Development and Validation of an Explainable Deep Learning Model to Predict In-Hospital Mortality for Patients With Acute Myocardial Infarction: Algorithm Development and Validation Study

Development and Validation of an Explainable Deep Learning Model to Predict In-Hospital Mortality for Patients With Acute Myocardial Infarction: Algorithm Development and Validation Study

Development and Validation of an Explainable Deep Learning Model to Predict In-Hospital Mortality for Patients With Acute Myocardial Infarction: Algorithm Development and Validation Study

Authors of this article:

Puguang Xie1 Author Orcid Image ;   Hao Wang1 Author Orcid Image ;   Jun Xiao1 Author Orcid Image ;   Fan Xu1 Author Orcid Image ;   Jingyang Liu1 Author Orcid Image ;   Zihang Chen2 Author Orcid Image ;   Weijie Zhao2 Author Orcid Image ;   Siyu Hou3 Author Orcid Image ;   Dongdong Wu4 Author Orcid Image ;   Yu Ma1 Author Orcid Image ;   Jingjing Xiao3 Author Orcid Image
Article Authors Cited by Tweetations Metrics

    Original Paper

    1Chongqing Emergency Medical Centre, Chongqing University Central Hospital, School of Medicine, Chongqing University, Chongqing, China

    2Bioengineering College, Chongqing University, Chongqing, China

    3Bio-Med Informatics Research Centre & Clinical Research Centre, Xinqiao Hospital, Army Medical University, Chongqing, China

    4Medical Big Data Research Centre, Chinese People's Liberation Army General Hospital, Beijing, China

    *these authors contributed equally

    Corresponding Author:

    Jingjing Xiao, PhD

    Bio-Med Informatics Research Centre & Clinical Research Centre, Xinqiao Hospital, Army Medical University

    No. 183 Xinqiao Street, Shapingba District

    Chongqing, 400037

    China

    Phone: 86 18502299862

    Email: shine636363@sina.com


    Background: Acute myocardial infarction (AMI) is one of the most severe cardiovascular diseases and is associated with a high risk of in-hospital mortality. However, the current deep learning models for in-hospital mortality prediction lack interpretability.

    Objective: This study aims to establish an explainable deep learning model to provide individualized in-hospital mortality prediction and risk factor assessment for patients with AMI.

    Methods: In this retrospective multicenter study, we used data for consecutive patients hospitalized with AMI from the Chongqing University Central Hospital between July 2016 and December 2022 and the Electronic Intensive Care Unit Collaborative Research Database. These patients were randomly divided into training (7668/10,955, 70%) and internal test (3287/10,955, 30%) data sets. In addition, data of patients with AMI from the Medical Information Mart for Intensive Care database were used for external validation. Deep learning models were used to predict in-hospital mortality in patients with AMI, and they were compared with linear and tree-based models. The Shapley Additive Explanations method was used to explain the model with the highest area under the receiver operating characteristic curve in both the internal test and external validation data sets to quantify and visualize the features that drive predictions.

    Results: A total of 10,955 patients with AMI who were admitted to Chongqing University Central Hospital or included in the Electronic Intensive Care Unit Collaborative Research Database were randomly divided into a training data set of 7668 (70%) patients and an internal test data set of 3287 (30%) patients. A total of 9355 patients from the Medical Information Mart for Intensive Care database were included for independent external validation. In-hospital mortality occurred in 8.74% (670/7668), 8.73% (287/3287), and 9.12% (853/9355) of the patients in the training, internal test, and external validation cohorts, respectively. The Self-Attention and Intersample Attention Transformer model performed best in both the internal test data set and the external validation data set among the 9 prediction models, with the highest area under the receiver operating characteristic curve of 0.86 (95% CI 0.84-0.88) and 0.85 (95% CI 0.84-0.87), respectively. Older age, high heart rate, and low body temperature were the 3 most important predictors of increased mortality, according to the explanations of the Self-Attention and Intersample Attention Transformer model.

    Conclusions: The explainable deep learning model that we developed could provide estimates of mortality and visual contribution of the features to the prediction for a patient with AMI. The explanations suggested that older age, unstable vital signs, and metabolic disorders may increase the risk of mortality in patients with AMI.

    J Med Internet Res 2024;26:e49848

    doi:10.2196/49848

    Keywords

    acute myocardial infarction; mortality; deep learning; explainable model; prediction 


    Background

    Acute myocardial infarction (AMI) refers to myocardial ischemia caused by acute blockage of coronary arteries, which is one of the most common cardiovascular emergencies [1]. AMI remains a leading cause of mortality worldwide, and patients with AMI have an estimated in-hospital mortality of 7% to 14% [2,3]. Therefore, the prediction of in-hospital mortality and assessment of risk factors for patients with AMI is of great interest to clinicians. Predicting the risk of mortality and conducting an individualized analysis of the risk factors in patients with AMI at initial diagnosis is critical for developing targeted interventions and improving the prognosis [4].

    A substantial effort has been devoted to developing models for predicting mortality in patients with AMI [5]. In the past, generalized linear models (eg, the logistic regression [LR] model and Cox proportional hazard model) have been used to predict mortality in patients with AMI [6-9]. However, the generalized linear models fail to capture the nonlinear relations of the massive, high-dimensional, and incomplete medical data, which hinder the practical use and further clinical application of the models [10]. With the rapid development of artificial intelligence, tree-based machine learning models, such as random forest (RF) [11] and extreme gradient boosting (XGBoost) [12], were used to address problems related to the prediction of mortality and achieved promising performance [13-15]. Nevertheless, tree-based models tend to be ineffective in the face of web-based learning and sparse classification features. More importantly, tree-based models failed to develop the model with powerful deep learning modules through the backpropagation method, which limits the further improvement of the model performance [16]. Deep learning is the most advanced branch of artificial intelligence, which can automatically learn feature representations from training data and provide new prospects for solving the existing limitations in tree-based models [17]. However, state-of-the-art deep learning models specific to tabular data have not been applied to the prediction of mortality in patients with AMI. In addition, while a complex model may perform better than a simpler one, it could lead to black-box problems, meaning that only the inputs and outputs of the model can be seen and it is difficult to understand how the inputs affect the predictions, which limit their clinical acceptance and raise ethical and legal questions [18]. The proposal of methods that could provide explanations for black-box models might facilitate clinicians in understanding predictions and making faster and more accurate treatment decisions [9,19,20]. At present, there is a lack of tools to predict in-hospital mortality of patients with AMI based on tabular data–specific deep learning models, and the potential of state-of-the-art deep learning models for predicting mortality in AMI is unclear. In addition, most mortality prediction models for patients with AMI proposed in previous studies are black-box models, which makes it difficult for clinicians to understand the predictions and use the models in clinical practice.

    Contribution

    In this study, we chose 3 state-of-the-art deep learning models to predict in-hospital mortality of patients with AMI and compared their performance with linear and tree-based models. In addition, we used an algorithm to interpret the best-performing black-box prediction model to obtain a tool for individualized mortality prediction and risk factor assessment in patients with AMI.


    Data Source

    This retrospective study included consecutive patients hospitalized with AMI from Chongqing University Central Hospital between July 2016 and December 2022, the Medical Information Mart for Intensive Care-IV (MIMIC-IV; version 2.2) database [21], and the Electronic Intensive Care Unit Collaborative Research Database (eICU-CRD; version 2.0) [22]. Chongqing University Central Hospital is a city-designated emergency tertiary medical center in China with a permanent population of >32 million. The MIMIC-IV database and eICU-CRD are 2 large free public databases. The MIMIC-IV database contains case information from 299,712 patients admitted to the intensive care unit or the emergency department at the Beth Israel Deaconess Medical Center from 2008 to 2019. The eICU-CRD included information from 139,367 patients from 208 hospitals in the United States between 2014 and 2015.

    Ethical Considerations

    Access to the MIMIC-IV database and eICU-CRD was approved by the Massachusetts Institute of Technology (Cambridge, Massachusetts) and Beth Israel Deaconess Medical Center (Boston, Massachusetts), and consent was obtained for the data collection. These databases are public, all patient data are anonymized, and data extracted from the databases do not require individual informed consent. PX attended a series of courses offered by the National Institutes of Health and was granted access to these databases after passing the required assessment (record ID 51524821). The study was conducted following the Declaration of Helsinki. Patient data obtained from the Chongqing University Central Hospital were also anonymized. Ethics approval was obtained from the Chongqing University Central Hospital Ethics Committee, and informed consent was waived.

    Population

    All patients diagnosed with AMI according to the ICD (International Classification of Diseases) in the Chongqing University Central Hospital between July 2016 and December 2022, MIMIC-IV database and eICU-CRD were included. ICD-9 (International Classification of Diseases, Ninth Revision) code 410 and ICD-10 (International Statistical Classification of Diseases, Tenth Revision) code I21 were used to identify the patients with AMI. For patients with AMI admitted multiple times during the study period, only the first admission was included. Exclusion criteria were (1) patients aged <18 years, (2) patients with unclear clinical outcome, and (3) patients with >20% of missing data (Figure 1).

    ‎
    Figure 1. Study profile: (A) Inclusion and exclusion criteria for the training and internal test data set; and (B) Inclusion and exclusion criteria for the external validation data set. AMI: acute myocardial infarction; eICU-CRD: Electronic Intensive Care Unit Collaborative Research Database; MIMIC-IV: Medical Information Mart for Intensive Care-IV.

    Data Extraction

    Patient data at Chongqing University Central Hospital were collected by reviewing electronic medical records. The data in the MIMIC-IV database and eICU-CRD were obtained using SQL. A total of 39 variables were incorporated into this study to determine the prediction models, including demographics, vital signs, laboratory parameters, comorbidities and drug use data based on previous studies, and clinical relevance and experience. All vital signs and laboratory data were the first measurements after admission. Only variables with <30% of missing values across the entire data set were considered for inclusion. This inclusion criterion, unlike the exclusion criterion for patients, was applicable to variables in the entire database. The predetermined threshold was selected according to the evidence from previous studies and experience [23,24]. The missing values were filled by the K-nearest neighbors (KNN) algorithm, and the hyperparameter K for KNN was determined using a 5-fold cross-validation in the training data set. The KNN model trained in the training data set was then used to fill in the missing values in the internal test and external validation data sets. In total, 24,004 missing data points were successfully filled by the KNN model. The clinical outcome assessed in this study was all-cause in-hospital mortality.

    Statistical Analysis

    Patients with AMI in the training data set were divided into a survival group and a nonsurvival group according to the clinical outcome. Shapiro-Wilk test was applied for the normality test. Continuous variables were presented as medians (IQRs), and categorical variables were presented as numbers and percentages. To test the differences within the groups, the 2-tailed t test and Kruskal-Wallis test were used for normal and nonnormal continuous variables, respectively. For categorical variables, between-group differences were evaluated using the chi-square test. P value <.05 was regarded as statistically significant.

    Model Development

    The data of the Chongqing University Central Hospital and eICU-CRD were merged and shuffled. The shuffled data set was then randomly divided into the training data set and internal test data set in a 7:3 ratio. The reason for this combination was that the data of the eICU-CRD were collected from 208 hospitals in the United States, and the generalization ability of the model may be improved by combining the data of eICU-CRD with the data from Chongqing University Central Hospital into a larger multicenter training data set. Patient data in the MIMIC-IV database were used for an independent external validation, which is independent of the Chongqing University Central Hospital and eICU-CRD data.

    Data preprocessing tasks were performed before training the prediction models. The KNN algorithm and z score were used to fill in missing values and normalize continuous variables, respectively.

    A total of 9 models were used to predict in-hospital mortality in patients with AMI, including state-of-the-art deep learning models (eg, Self-Attention and Intersample Attention Transformer [SAINT] [25], TabTransformer [26], and TabNet [27]), a basic deep learning model (eg, multilayer perceptron [MLP] [28]), tree-based machine learning models (eg, RF [11], XGBoost [12], light gradient boosting machine [29], and categorical boosting [30]), and a linear model (LR [31]). An optimization procedure based on nested 5-fold cross-validation Bayesian optimization was used for each model with 100 iterations to determine the hyperparameters. Additional details on hyperparameters setting are provided in Tables S1-S8 in Multimedia Appendix 1. The complete training data set was then used to train the final models with the optimal hyperparameters. Code for model development is publicly available [32].

    Model Evaluation

    The area under the receiver operating characteristic curve (AUC) was used to assess the performance of the prediction models in the internal test data set and external validation data set. The cutoff thresholds were determined according to the Youden index, and 5 metrics were further calculated, including balanced accuracy, sensitivity, specificity, precision, and F1-score. CIs were calculated using 1000 bootstrap samples. The calibration curve and Brier score were used to evaluate the reliability of probability estimates of the prediction models; decision curve analysis was used to evaluate the clinical utility of the models.

    Model Explanation

    The Shapley Additive Explanations (SHAP) algorithm was used to interpret the predictions to obtain global and local influences of the relevant features on the predictions. With a set of feature values of a patient, a SHAP value represents how much a single feature’s value influences the difference between the actual prediction and the average prediction of the model in interaction with other features. Briefly, the mean prediction of the model plus the sum of the SHAP values of a sample is equal to the prediction of that sample [33,34].

    We first used the SHAP algorithm to visualize the global relative importance of features to the predictions. The rationality of the abovementioned results was then verified by statistical methods. Patients in the training data set were further divided into a survival group and a nonsurvival group. Statistical test was used to evaluate the differences in features between the groups, and the test results were compared with those obtained by the SHAP algorithm. Finally, we linearly mapped the SHAP value to the probability of driving an increase or decrease in predicted mortality, quantifying and visualizing the contribution of each feature to the prediction in individuals.


    Statistical Analysis Results

    A total of 2428 consecutive patients with AMI who were admitted to Chongqing University Central Hospital between July 2016 and December 2022 and 8527 patients with AMI from the eICU-CRD were included in the analysis. A total of 10,955 patients were randomly divided into a training data set of 7668 (70%) patients and an internal test data set of 3287 (30%) patients. The median age of patients in the training and internal test data sets was 66 (IQR 56-76) years, with 65.6% (5030/7668) and 68.09% (2238/3287) of male participants, respectively. In total, 8.74% (670/7668) of the patients each in the training data set and 8.73% (287/3287) of the patients in the internal test data set died during their hospital stay (Table 1).

    A total of 9355 patients with AMI from the MIMIC-IV database were included for independent external validation. The median age of patients in the external validation group was 72 (IQR 62-82) years, higher than that in the training or internal test data set. In total, 61.17% (5722/9355) of the patients were male participants, and 9.12% (853/9355) of the patients died during the hospitalization (Table 1).

    Table 1. Baseline characteristics of patients with acute myocardial infarction included in the models (N=10,955).
    VariablesTraining data set (n=7668)Internal test data set (n=3287)External validation data set (n=9355)
    Demographic data

    		Age (years), median (IQR)66 (56-76)66 (56-76)72 (62-82)

    		Female participants, n (%)2638 (34.4)1049 (31.9)3633 (38.8)
    Vital signs,median (IQR)

    		Heart rate (beats/minute)81 (70-96)81 (70-95)83 (72-96)

    		Respiratory rate (minute−1)20 (17-21)20 (17-21)18 (16-20)

    		Systolic blood pressure (mm Hg)129 (111-149)129 (112-148)126 (109-145)

    		Diastolic blood pressure (mm Hg)76 (64-88)76 (64-88)69 (58-82)

    		Temperature (°C)36.6 (36.4-36.8)36.6 (36.4-36.8)36.7 (36.4-37)
    Laboratory data, median (IQR)

    		Troponin T or I (ng/mL)0.74 (0.10-4.85)0.69 (0.10-4.93)0.32 (0.11-1.01)

    		White blood cell (k/µ)10.6 (8.2-13.6)10.7 (8.3-13.7)9.30 (7.1-12.4)

    		Red blood cell (m/µ)4.42 (3.89-4.87)4.43 (3.91-4.87)3.88 (3.34-4.41)

    		Hemoglobin (g/dL)13.5 (11.7-14.8)13.4 (11.8-14.9)11.6 (9.9-13.2)

    		Hematocrit (%)40.1 (35.4-43.9)40.2 (35.7-44)35.2 (30.5-39.6)

    		Platelet (k/µ)215 (173-265)217 (174-268)208 (164-262)

    		Prothrombin time (seconds)13.3 (12.3-14.5)13.3 (12.3-14.5)12.9 (11.8-14.7)

    		International normalized ratio1.06 (1-1.2)1.05 (1-1.2)1.2 (1.1-1.3)

    		Sodium (mEq/L)138 (136-140)138 (136-140)139 (137-141)

    		Potassium (mEq/L)4.00 (3.70-4.37)4 (3.7-4.4)4.2 (3.8-4.5)

    		Chloride (mEq/L)103.3 (100-106)103.4 (100-106)103 (100-106)

    		Calcium (mg/dL)8.90 (8.5-9.3)8.9 (8.41-9.3)8.7 (8.3-9.1)

    		Magnesium (mg/dL)1.99 (1.8-2.2)1.99 (1.77-2.19)2 (1.8-2.2)

    		Creatinine (mg/dL)1.04 (0.84-1.38)1.03 (0.82-1.40)1.1 (0.8-1.6)

    		Blood urea nitrogen (mg/dL)19 (14- 28)19 (14-28)21 (15-35)
    Comorbidities, n (%)

    		Hypertension4616 (60.2)1965 (59.8)3962 (42.4)

    		Hyperlipidemia820 (10.7)397 (12.1)4940 (52.8)

    		Diabetes2516 (32.8)1087 (33.1)3813 (40.8)

    		Heart failure1987 (25.9)854 (26)4298 (45.9)

    		Renal failure1648 (21.5)760 (23.1)4446 (47.5)

    		Cerebral vascular disease788 (10.3)348 (10.6)1106 (11.8)

    		Peripheral vascular disease1062 (13.8)439 (13.4)1324 (14.2)

    		Chronic obstructive pulmonary disease864 (11.3)386 (11.7)688 (7.4)
    Operations, n (%)

    		Coronary arteriography3855 (50.3)1663 (50.6)2839 (30)

    		Percutaneous coronary intervention3083 (40.2)1359 (41.3)2954 (31.6)
    Drugs, n (%)

    		Aspirin4941 (64.4)2126 (64.7)8534 (91.2)

    		Clopidogrel2070 (27)872 (26.5)4385 (46.9)

    		Ticagrelor1491 (19.4)629 (19.1)622 (6.6)

    		Angiotensin-converting enzyme inhibitor1827 (23.8)765 (23.3)4001 (42.8)

    		Angiotensin receptor blockers240 (3.1)118 (3.6)1158 (12.4)

    		Statins3728 (48.6)1596 (48.6)8276 (88.5)

    		β-Blockers3565 (46.5)1555 (47.3)7625 (81.5)
    Clinical outcome

    		All-cause in-hospital mortality, n (%)670 (8.7)287 (8.7)853 (9.1)

    		Length of hospital stay (days), median (IQR)4.98 (2.50-9.25)4.95 (2.42-9.51)5.33 (2.80-9.85)

    Model Performance

    The AUCs of 9 models in the internal test data set and external validation data set are shown in Figures 2A and 2B (AUC=1 indicates perfect prediction and AUC=0.5 indicates random prediction), and the other 5 evaluation metrics, including balanced accuracy, sensitivity, specificity, precision, and F1-score are listed in Tables 2 and 3.

    ‎
    Figure 2. Performance of the models in predicting in-hospital mortality of patients with acute myocardial infarction in the internal test set and external validation set: (A) and (B), AUCs of the models; (C) and (D), calibration plots of the models; and (E) and (F), decision curves of the models. AUC: area under the receiver operating characteristic curve; CatBoost: categorical boosting; LightGBM: light gradient boosting machine; LR: logistic regression; MLP: multilayer perceptron; RF: random forest; SAINT: Self-Attention and Intersample Attention Transformer; XGBoost: extreme gradient boosting.
    Table 2. Values of the evaluation metrics of the models in the internal test data set.
    ModelsBalanced accuracy (95% CI)Sensitivity (95% CI)Specificity (95% CI)Precision (95% CI)F1-score (95% CI)AUCa (95% CI)
    Logistic regression0.76 (0.74-0.78)0.85 (0.81-0.89)0.67 (0.65-0.68)0.20 (0.17-0.22)0.32 (0.29-0.35)0.83 (0.81-0.85)
    Random forest0.72 (0.69-0.75)0.58 (0.52-0.64)0.87 (0.86-0.88)0.30 (0.26-0.34)0.39 (0.35-0.43)0.85 (0.82-0.87)
    Extreme gradient boosting0.75 (0.73-0.78)0.74 (0.69-0.78)0.77 (0.75-0.78)0.23 (0.21-0.26)0.35 (0.32-0.39)0.85 (0.83-0.87)
    Light gradient boosting machine0.71 (0.68-0.73)0.52 (0.47-0.58)0.89 (0.88-0.91)0.32 (0.28-0.36)0.40 (0.35-0.44)0.85 (0.83-0.87)
    Categorical boosting0.76 (0.73-0.78)0.74 (0.69-0.79)0.78 (0.76-0.80)0.24 (0.21-0.27)0.37 (0.33-0.40)0.85 (0.83-0.87)
    Multilayer perceptron0.77 (0.74-0.79)0.79 (0.75-0.84)0.74 (0.73-0.76)0.23 (0.20-0.25)0.32 (0.35-0.39)0.84 (0.82-0.86)
    TabNet0.76 (0.73-0.78)0.86 (0.82-0.90)0.65 (0.64-0.67)0.19 (0.17-0.21)0.31 (0.28-0.34)0.81 (0.79-0.84)
    TabTransformer0.77 (0.75-0.79)0.85 (0.81-0.89)0.69 (0.67-0.71)0.21 (0.18-0.23)0.33 (0.30-0.37)0.84 (0.82-0.86)
    Self-Attention and Intersample Attention Transformer0.76 (0.73-0.78)0.74 (0.69-0.79)0.77 (0.75-0.78)0.23 (0.21-0.26)0.36 (0.32-0.39)0.86 (0.84-0.88)

    aAUC: area under the receiver operating characteristic curve.

    Table 3. Values of the evaluation metrics of the models in the external validation data set.
    ModelsBalanced accuracy (95% CI)Sensitivity (95% CI)Specificity (95% CI)Precision (95% CI)F1-score (95% CI)AUCa (95% CI)
    Logistic regression0.73 (0.74-0.76)0.91 (0.89-0.93)0.58 (0.57-0.59)0.18 (0.17-0.19)0.30 (0.28-0.31)0.83 (0.82-0.85)
    Random forest0.75 (0.73-0.77)0.70 (0.67-0.73)0.80 (0.79-0.81)0.26 (0.24-0.28)0.38 (0.36-0.40)0.84 (0.82-0.85)
    Extreme gradient boosting0.75 (0.74-0.77)0.82 (0.79-0.84)0.69 (0.68-0.70)0.21 (0.19-0.22)0.33 (0.31-0.35)0.84 (0.83-0.85)
    Light gradient boosting machine0.73 (0.71-0.75)0.57 (0.54-0.61)0.89 (0.88-0.89)0.34 (0.31-0.36)0.42 (0.40-0.45)0.85 (0.85-0.86)
    Categorical boosting0.76 (0.75-0.77)0.81 (0.78-0.84)0.71 (0.70-0.72)0.22 (0.21-0.23)0.35 (0.33-0.36)0.85 (0.83-0.86)
    Multilayer perceptron0.75 (0.74-0.76)0.89 (0.86-0.91)0.62 (0.61-0.63)0.19 (0.18-0.20)0.31 (0.29-0.33)0.84 (0.82-0.85)
    TabNet0.69 (0.68-0.70)0.93 (0.91-0.94)0.45 (0.44-0.46)0.14 (0.14-0.15)0.25 (0.24-0.26)0.78 (0.77-0.80)
    TabTransformer0.75 (0.74-0.76)0.92 (0.90-0.94)0.57 (0.57-0.59)0.18 (0.17-0.19)0.30 (0.28-0.31)0.83 (0.82-0.85)
    Self-Attention and Intersample Attention Transformer0.77 (0.76-0.78)0.87 (0.85-0.90)0.67 (0.66-0.68)0.21 (0.20-0.22)0.34 (0.32-0.36)0.85 (0.84-0.87)

    aAUC: area under the receiver operating characteristic curve.

    In general, the SAINT model performed best in both the internal test data set and the external validation data set among the 9 prediction models, with an AUC of 0.86 (95% CI 0.84-0.88) and 0.85 (95% CI 0.84-0.87), respectively. The corresponding balanced accuracy of the internal test data set and external validation data set were 0.76 (95% CI 0.73-0.78) and 0.77 (95% CI 0.76-0.78), sensitivities were 0.74 (95% CI 0.69-0.79) and 0.87 (95% CI 0.85-0.90), specificities were 0.77 (95% CI 0.75-0.78) and 0.67 (95% CI 0.66-0.68), precision values were 0.23 (95% CI 0.21-0.26) and 0.21 (95% CI 0.20-0.22), and F1-score were 0.36 (95% CI 0.32-0.39) and 0.34 (95% CI 0.32-0.36), respectively. The calibration curves demonstrated good consistency between the predicted and observed probabilities for the SAINT model with curves close to the 45° line, and the Brier scores were 0.07 (95% CI 0.01-0.21) and 0.07 (95% CI 0.02-0.17) in the internal test data set and external validation data set, respectively (Figures 2C and 2D). In addition, the decision curve analysis (Figures 2E and 2F) illustrated that the SAINT model was clinically beneficial in both the internal test data set and external validation data set.

    Moreover, we tried to train the model using different data sets and evaluated the performance of the models in different external validation data sets. The results showed that the SAINT model performed better than other models (Figures S2 and S3 and Tables S9-S12 in Multimedia Appendix 1). Furthermore, we evaluated the model’s predictive power using the measurements obtained within the first 2 hours of patient admission. This evaluation aimed to assess the predictive power of the models for the early diagnostic stage of patients with AMI. The results demonstrated that although the SAINT model still outperformed the other models, there was a clear decrease in the predictive power of all models, with an average decrease of 0.06 (SD 0.02) in the AUC. (Figure S4 and Table S13 in Multimedia Appendix 1).

    Model Explanation Results

    The relative importance and contribution of the top 20 features to the SAINT model are shown in Figure 3A, which was obtained by the SHAP algorithm interpreting the predictions of the SAINT model. The importance and contribution of the full features to the SAINT model are presented in Figure S1 in Multimedia Appendix 1. Age was ranked as the most important feature for the prediction model, followed by heart rate, temperature, white blood cell level, systolic blood pressure, magnesium level, prothrombin time, calcium level, creatinine level, and sodium level. Considering the impact of age on the predictions as an example, advanced age was commonly associated with a higher risk of mortality, and younger age generally drove the predictions toward survival (red dots represent a high value of the features, which on the right side of the x-axis represents an increased risk of mortality). Similar explanations can be applied to other features, and most interpretations were consistent with clinical experience. It is notable that features can drive predictions in different directions (increasing or decreasing mortality predictions) for patients with different features, unlike previous scoring systems (eg, Global Registry of Acute Coronary Events risk score [35] and Thrombolysis in Myocardial Infarction score [36]), where the influence of a given feature value on the prediction is invariant. The 9 most important features for the prediction models showed a statistically significant difference between the survival group and the nonsurvival group in the training data set (P<.05), which further supports the rationality of the explanations of the features (Table 4).

    We further mapped the SHAP values to the probability of driving an increase or decrease in mortality and visualized the application of the method to 1 deceased patient and 1 surviving patient in the external validation data set (Figures 3B and 3C). In the case of the deceased patient, the patient was female and aged 89 years with a history of heart failure and renal failure. The model predicted that the risk of mortality of the patients was 0.551. High levels of serum sodium (185 mEq/L) and serum magnesium (3.2 mg/dL) drove a 0.29 and 0.21 increase in the risk of mortality, respectively, whereas a high serum chloride level was associated with a 0.17 reduction in the risk of mortality. A similar explanation can be applied to other features. The prediction was driven by 39 features used for model training, and the sum of the SHAP values for all features plus the baseline risk equaled the predicted risk of mortality. The baseline risk score was obtained by calculating the average predicted risk of mortality among all patients in the training data set (0.079). Therefore, the SHAP algorithm made our model explainable at both the global and individual levels.

    ‎
    Figure 3. The impact of the input features on predictions: (A) the effect of the input features on the model at the global level. Each dot represents the influence of a feature on the prediction for a patient. Red dots indicate a higher value of the features, and blue dots indicate a lower value of the features. A dot on the left side of the x-axis represents that the feature value decreased mortality prediction, and a dot on the right side of the x-axis represents that the feature value increased mortality prediction, (B) an example of risk factor analysis for a patient in the external validation set (in-hospital death), and (C) an example of risk factor analysis for a patient in the external validation set (survival at hospital discharge). SHAP: Shapley Additive Explanations.
    Table 4. Baseline characteristics of patients with acute myocardial infarction in the training data set by clinical outcomes (n=7668).
    VariablesSurvival group (n=6998)Nonsurvival group (n=670)P valuea
    Demographic data

    		Age (years), median (IQR)66 (56-75)75 (66-83)<.001

    		Sex, n (%)<.001


    		Male4661 (66.6)369 (55.1)


    		Female2337 (33.4)301 (44.9)
    Vital signs, median (IQR)

    		Heart rate (beats/minute)80 (69-94)92 (78-108)<.001

    		Respiratory rate (minute−1)20 (17-21)20 (18-25)<.001

    		Systolic blood pressure (mm Hg)130 (113-149)116 (98-142)<.001

    		Diastolic blood pressure (mm Hg)77 (65-88)68 (57-81)<.001

    		Temperature (°C)36.6 (36.4-36.8)36.5 (36.1-36.8)<.001
    Laboratory data, median (IQR)

    		Troponin T or I (ng/mL)0.66 (0.10-4.47)1.77 (0.26-8.78)<.001

    		White blood cell (k/µ)10.4 (8.2-13.4)12.4 (9.1-16.8)<.001

    		Red blood cell (m/µ)4.45 (3.94-4.88)4.05 (3.48-4.59)<.001

    		Hemoglobin (g/dL)13.6 (11.8-14.9)12.1 (10.3-14)<.001

    		Hematocrit (%)40.4 (35.9-44)37 (32.1-42.4)<.001

    		Platelet (k/µ)215 (174-264)213 (153-270).06

    		Prothrombin time (seconds)13.2 (12.2-14.3)14.5 (13.2-17.2)<.001

    		International normalized ratio1.04 (1-1.16)1.2 (1.08-1.49)<.001

    		Sodium (mEq/L)138 (136-140)138 (135-141).06

    		Potassium (mEq/L)4 (3.7-4.3)4.2 (3.73-4.70)<.001

    		Chloride (mEq/L)103.6 (100.4-106)103 (98.5-106)<.001

    		Calcium (mg/dL)8.9 (8.5-9.3)8.6 (8.18-9.10)<.001

    		Magnesium (mg/dL)1.99 (1.80-2.19)2 (1.71-2.30)<.001

    		Creatinine (mg/dL)1.01 (0.82-1.30)1.42 (1.06-2.14)<.001

    		Blood urea nitrogen (mg/dL)19 (14-27)27 (18-41)<.001
    Comorbidities, n (%)

    		Hypertension4209 (60.1)407 (60.7).76

    		Hyperlipidemia767 (11)53 (7.9).02

    		Diabetes2249 (32.1)267 (39.9)<.001

    		Heart failure1688 (24.1)299 (44.6)<.001

    		Renal failure1333 (19)315 (47)<.001

    		Cerebral vascular disease670 (9.6)118 (17.6)<.001

    		Peripheral vascular disease971 (13.9)91 (13.6).83

    		Chronic obstructive pulmonary disease749 (10.7)115 (17.2)<.001
    Operations, n (%)

    		Coronary arteriography3670 (52.4)185 (27.6)<.001

    		Percutaneous coronary intervention2920 (41.7)163 (24.3)<.001
    Drugs, n (%)

    		Aspirin4583 (65.5)358 (53.4)<.001

    		Clopidogrel1943 (27.8)127 (19)<.001

    		Ticagrelor1413 (20.2)78 (11.6)<.001

    		Angiotensin-converting enzyme inhibitor1777 (25.4)50 (7.5)<.001

    		Angiotensin receptor blocker233 (3.3)7 (1).001

    		Statins3512 (50.2)216 (32.2)<.001

    		β-Blockers3369 (48.1)196 (29.3)<.001
    Clinical outcome, median (IQR)

    		Length of hospital stay (days)5.15 (2.59-9.34)3.71 (1.23-8.27)<.001

    aItalicized P values (<.05) were considered statistically significant.


    Principal Findings

    In this multicenter, retrospective study, we trained 3 state-of-the-art deep learning models on patients with AMI from Chongqing University Central Hospital and patient data from the eICU-CRD and compared their performance with 4 advanced tree-based models, an LR model, and an MLP. We performed both internal and external validation (MIMIC-IV database) to confirm the predictive ability of the models and ensure the reliability of the results. In addition, the SHAP algorithm was used to interpret the predictions of the best-performing model (SAINT) to quantify and visualize the impact of the features on the predicted mortality at global and individual levels.

    The models developed in this study had similar AUC in the internal test data set and the external validation data set, which demonstrated a good generalization ability of the models. The training data set we used was collected from 209 centers in Asia and the United States, and the multicenter data contributed to improving the performance of the model, especially in terms of generalization ability. However, although the specificities of the models in the internal test data set were generally higher and the sensitivities were lower than that in the external validation, the models tended to overestimate patient mortality in the external validation data set, leading to false positive results. The SAINT model is a hybrid deep learning–based approach specifically designed to solve tabular data problems and outperforms tree-based models over a variety of benchmark tasks [25]. In this study, the SAINT model performed best in both the internal test data set and the external validation data set among the 9 prediction models, with the highest AUC and lowest Brier score. In addition, the decision curve analysis indicated that the SAINT model could provide net benefits across an extensive range of threshold values. The SAINT model attained the highest balanced accuracy during evaluation on the external validation data set, indicating that the model had strong classification ability. Nevertheless, it is worth noting that both precision and specificity metrics associated with the SAINT model did not exhibit high values, implying a proclivity of the model to overestimate the mortality risk among patients with AMI. According to these results, state-of-the-art deep learning methods have greater potential in accurately predicting mortality in patients with AMI than linear, MLP, and tree-based models.

    The SHAP algorithm was used to obtain an interpretation of the predictions of the SAINT model, and the impact of most features on the predictions is consistent with clinical practice and previous evidence. For instance, advanced age, unstable vital signs, presence of infection, metabolic disorders, and coagulation disorders lead to increased predicted mortality. The results of the model interpretation indicated that advanced age was ranked as the most important risk factor for mortality. In addition, as patients age, their physical function and resistance decline, which may increase their risk of mortality [37,38]. Unstable vital signs (eg, elevated heart rate and respiratory rate and decreased levels of systolic blood pressure and temperature) usually indicate that the patient is in a critical condition, and such patients tend to have a worse prognosis [39,40]. Similarly, the presence of infection (eg, elevated levels of white blood cells) may impair physical function and reduce the effectiveness of treatment [41]. In addition, metabolic and coagulation disorders (eg, elevated levels of magnesium and sodium and increased prothrombin time) are associated with an increased risk of mortality [42].

    LR models and Cox proportional hazard models are commonly used tools for predicting the mortality of patients with AMI [6]. However, the rapid development of medical information technology brings massive, nonlinear, and high-dimensional data, which challenge the generalized linear model [43]. Tree-based machine learning models (eg, RF, XGBoost, and light gradient boosting machine) achieved substantial gains in the predictive accuracy of tabular data [4,13,14]. Nevertheless, tree-based models were developed directly using the information entropy of data without calculating the loss function and performing backpropagation, which limits the model from taking advantage of powerful deep learning modules and conducting web-based learning [29,30]. Recently, deep learning models have received extensive attention in the screening, diagnosis, and prognosis prediction of diseases, and their performance even outperformed that of clinicians in some aspects [44]. A recent, large population-based study by Mohammad et al [45] revealed the tremendous potential of deep learning models to identify the risk of adverse outcomes in patients with AMI. However, the prediction model used in this study is the most basic deep learning model (MLP), and the ability of state-of-the-art deep learning models to predict the mortality of patients with AMI remains unclear. In addition, the model used in the study by Mohammad et al [45] is regarded as a black-box model due to the complex internal calculation process, that is, we can only see the input and final predictions of the model but cannot understand how the model makes decisions. We used the SHAP algorithm to quantify and visualize the features that drive the predictions, which combines the advantages of complex deep learning models and simple linear models to solve the problem of insufficient predictive ability of the linear model and black-box model of the complex deep learning model. The effect of each feature on the prediction is not isolated but derived from the interaction with the other features, which makes it superior to the fixed feature weight in the linear models [33]. Furthermore, the SHAP algorithm not only provides the impact of the features on the model but also provides the interpretation at an individual level. Such a patient-level interpretable model allows clinicians to combine the predictions and explanations with their empirical knowledge to facilitate decision-making [19].

    Limitations

    This study has several limitations. First, to enable the models to learn more comprehensive information and improve performance, up to 39 features were used to develop the model. Nevertheless, the variables included in the study were limited by the retrospective design and the use of publicly available databases. Second, there were missing values in the raw data; although the KNN algorithm was used to complete them, the real data could provide more valuable information. This limitation was particularly significant during the early stages of patient hospital admission when the available predictors were limited. Even with the application of the KNN algorithm to complete the missing values, the predictive power of the model could be compromised. Third, although the SHAP algorithm was used to obtain the explanations of the models, the impact of features on the predictions is not equal to the association in the causal chain. In addition, features such as age and sex, while influencing predictions, were not subject to human intervention. However, the explanations of the relationship between features and predictions may guide us in examining causation.

    Conclusions

    In conclusion, in this multicenter study, we developed and externally validated 9 prediction models for in-hospital mortality in patients with AMI. Our results indicate that state-of-the-art deep learning models may provide a promising alternative solution to further improve the predictive ability of mortality in patients with AMI. The explanations of the model suggest that older age, unstable vital signs, and metabolic disorders may increase the risk of mortality in patients with AMI. Transparent and explainable models are expected to support clinical practice, trial design, and future research. However, future prospective studies are needed to confirm the findings.

    Acknowledgments

    This study was supported by Independent Research Project of Medical Engineering Laboratory of Chinese PLA General Hospital (2022SYSZZKY07), National Natural Science Foundation of China (62076247, 72374206), National Key R&D Program of China (2021ZD0140408), and Technology Innovation and Application Development Special Project of Chongqing, China (CSTB2023TIAD-KPX0047).

    Data Availability

    The data sets from the Medical Information Mart for Intensive Care-IV database and Electronic Intensive Care Unit Collaborative Research Database presented in this study can be found in web-based repositories [46,47]. The data set from Chongqing University Central Hospital is not publicly available but is available from the corresponding author on reasonable request.

    Authors' Contributions

    PX was responsible for the study design, data acquisition, analysis, and interpretation and for preparing the first draft of the manuscript. HW, JX, FX, JL, ZC, WZ, SH, and DW were responsible for recruiting study participants and providing clinical input to the study design. YM and JX designed the study and helped write the manuscript. All authors had an opportunity to review the manuscript and approved its final submitted version.

    Conflicts of Interest

    None declared.

    Multimedia Appendix 1

    Supplementary materials.

    DOCX File , 1018 KB
    1. Reed GW, Rossi JE, Cannon CP. Acute myocardial infarction. Lancet. Jan 2017;389(10065):197-210. [CrossRef]
    2. Cui Y, Hao K, Takahashi J, Miyata S, Shindo T, Nishimiya K, et al. Age-specific trends in the incidence and in-hospital mortality of acute myocardial infarction over 30 years in Japan— report from the Miyagi AMI registry study. Circ J. 2017;81(4):520-528. [CrossRef]
    3. Szummer K, Wallentin L, Lindhagen L, Alfredsson J, Erlinge D, Held C, et al. Improved outcomes in patients with ST-elevation myocardial infarction during the last 20 years are related to implementation of evidence-based treatments: experiences from the SWEDEHEART registry 1995-2014. Eur Heart J. Nov 01, 2017;38(41):3056-3065. [CrossRef] [Medline]
    4. Figtree GA, Vernon ST, Hadziosmanovic N, Sundström J, Alfredsson J, Arnott C, et al. Mortality in STEMI patients without standard modifiable risk factors: a sex-disaggregated analysis of SWEDEHEART registry data. Lancet. Mar 20, 2021;397(10279):1085-1094. [CrossRef] [Medline]
    5. Castro-Dominguez Y, Dharmarajan K, McNamara RL. Predicting death after acute myocardial infarction. Trends Cardiovasc Med. Feb 2018;28(2):102-109. [CrossRef] [Medline]
    6. Funk M, Pooley-Richards RL. Predicting hospital mortality in patients with acute myocardial infarction. Am J Crit Care. May 1994;3(3):168-176. [Medline]
    7. Austin PC, Tu JV. Automated variable selection methods for logistic regression produced unstable models for predicting acute myocardial infarction mortality. J Clin Epidemiol. Nov 2004;57(11):1138-1146. [CrossRef] [Medline]
    8. Amin AP, Spertus JA, Reid KJ, Lan X, Buchanan DM, Decker C, et al. The prognostic importance of worsening renal function during an acute myocardial infarction on long-term mortality. Am Heart J. Dec 2010;160(6):1065-1071. [CrossRef] [Medline]
    9. Huang Y, Zheng Z, Ma M, Xin X, Liu H, Fei X, et al. Improving the performance of outcome prediction for inpatients with acute myocardial infarction based on embedding representation learned from electronic medical records: development and validation study. J Med Internet Res. Aug 03, 2022;24(8):e37486. [CrossRef] [Medline]
    10. Mood C. Logistic regression: why we cannot do what we think we can do, and what we can do about it. Eur Sociol Rev. Mar 09, 2010;26(1):67-82. [CrossRef]
    11. Breiman L. Random forests. Mach Learn. 2001;45:5-32. [FREE Full text]
    12. 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. Presented at: KDD '16; August 13-17, 2016;785-794; San Francisco, CA. URL: https://dl.acm.org/doi/10.1145/2939672.2939785 [CrossRef]
    13. Khera R, Haimovich J, Hurley NC, McNamara R, Spertus JA, Desai N, et al. Use of machine learning models to predict death after acute myocardial infarction. JAMA Cardiol. Jun 01, 2021;6(6):633-641. [CrossRef] [Medline]
    14. Oliveira M, Seringa J, Pinto FJ, Henriques R, Magalhães T. Machine learning prediction of mortality in acute myocardial infarction. BMC Med Inform Decis Mak. Apr 18, 2023;23(1):70. [CrossRef] [Medline]
    15. Xie P, Yang C, Yang G, Jiang Y, He M, Jiang X, et al. Mortality prediction in patients with hyperglycaemic crisis using explainable machine learning: a prospective, multicentre study based on tertiary hospitals. Diabetol Metab Syndr. Mar 11, 2023;15(1):44. [CrossRef] [Medline]
    16. Borisov V, Leemann T, Seßler K, Haug J, Pawelczyk M, Kasneci G. Deep neural networks and tabular data: a survey. IEEE Trans Neural Netw Learn Syst. 2022.:1-21. [FREE Full text] [CrossRef]
    17. Shrestha A, Mahmood A. Review of deep learning algorithms and architectures. IEEE Access. 2019;7:53040-53065. [CrossRef]
    18. Yang G, Ye Q, Xia J. Unbox the black-box for the medical explainable AI via multi-modal and multi-centre data fusion: a mini-review, two showcases and beyond. Inf Fusion. Jan 2022;77:29-52. [CrossRef] [Medline]
    19. Liang Y, Li S, Yan C, Li M, Jiang C. Explaining the black-box model: a survey of local interpretation methods for deep neural networks. Neurocomputing. Jan 2021;419:168-182. [CrossRef]
    20. Li J, Liu S, Hu Y, Zhu L, Mao Y, Liu J. Predicting mortality in intensive care unit patients with heart failure using an interpretable machine learning model: retrospective cohort study. J Med Internet Res. Aug 09, 2022;24(8):e38082. [CrossRef] [Medline]
    21. Johnson AE, Bulgarelli L, Shen L, Gayles A, Shammout A, Horng S, et al. Author correction: MIMIC-IV, a freely accessible electronic health record dataset. Sci Data. Jan 16, 2023;10(1):31. [CrossRef] [Medline]
    22. Pollard TJ, Johnson AE, Raffa JD, Celi LA, Mark RG, Badawi O. The eICU collaborative research database, a freely available multi-center database for critical care research. Sci Data. Sep 11, 2018;5(1):180178. [CrossRef] [Medline]
    23. Payrovnaziri SN, Xing A, Salman S, Liu X, Bian J, He Z. Assessing the impact of imputation on the interpretations of prediction models: a case study on mortality prediction for patients with acute myocardial infarction. AMIA Jt Summits Transl Sci Proc. 2021;2021:465-474. [Medline]
    24. Emmanuel T, Maupong T, Mpoeleng D, Semong T, Mphago B, Tabona O. A survey on missing data in machine learning. J Big Data. Oct 27, 2021;8(1):140. [CrossRef] [Medline]
    25. Somepalli G, Goldblum M, Schwarzschild A, Bruss CB, Goldstein T. Saint: Improved neural networks for tabular data via row attention and contrastive pre-training. Preprint posted online June 2, 2021.
    26. Huang X, Khetan A, Cvitkovic M, Karnin Z. TabTransformer: tabular data modeling using contextual embeddings. Preprint posted online December 11, 2020. [FREE Full text]
    27. Arik S, Pfister T. TabNet: attentive interpretable tabular learning. AAAI Conf Artif Intell. 2021;35(8):6679-6687. [FREE Full text] [CrossRef]
    28. McCulloch WS, Pitts W. A logical calculus of the ideas immanent in nervous activity. Bull Math Biophys. Dec 1943;5(4):115-133. [CrossRef]
    29. Ke G, Meng Q, Finley T, Wang T, Chen W, Ma W, et al. Lightgbm: a highly efficient gradient boosting decision tree. In: Proceedings of the 31st International Conference on Neural Information Processing Systems. 2017. Presented at: NIPS '17; December 4-9, 2017;3149-3157; Long Beach, CA. URL: https://dl.acm.org/doi/10.5555/3294996.3295074
    30. Prokhorenkova L, Gusev G, Vorobev A, Dorogush AV, Gulin A. CatBoost: unbiased boosting with categorical features. In: Proceedings of the 32nd International Conference on Neural Information Processing Systems. 2018. Presented at: NIPS '18; December 3-8, 2018;6639-6649; Montréal, QC. URL: https://dl.acm.org/doi/abs/10.5555/3327757.3327770
    31. DeMaris A. A tutorial in logistic regression. J Marriage Fam. Nov 1995;57(4):956. [CrossRef]
    32. PuguangXie/AMI. GitHub. URL: https://github.com/PuguangXie/AMI [accessed 2024-04-29]
    33. Lundberg SM, Erion G, Chen H, DeGrave A, Prutkin JM, Nair B, et al. From local explanations to global understanding with explainable AI for trees. Nat Mach Intell. Jan 17, 2020;2(1):56-67. [CrossRef] [Medline]
    34. Lundberg SM, Lee SI. A unified approach to interpreting model predictions. In: Proceedings of the 31st International Conference on Neural Information Processing Systems. 2017. Presented at: NIPS '17; December 4-9, 2017;4768-4777; Long Beach, CA. URL: https://dl.acm.org/doi/10.5555/3295222.3295230
    35. Granger CB, Goldberg RJ, Dabbous O, Pieper KS, Eagle KA, Cannon CP, et al. Global Registry of Acute Coronary Events Investigators. Predictors of hospital mortality in the global registry of acute coronary events. Arch Intern Med. Oct 27, 2003;163(19):2345-2353. [CrossRef] [Medline]
    36. Morrow DA, Antman EM, Charlesworth A, Cairns R, Murphy SA, de Lemos JA, et al. TIMI risk score for ST-elevation myocardial infarction: a convenient, bedside, clinical score for risk assessment at presentation: An intravenous nPA for treatment of infarcting myocardium early II trial substudy. Circulation. Oct 24, 2000;102(17):2031-2037. [CrossRef] [Medline]
    37. Cosentino N, Resta ML, Somaschini A, Campodonico J, Lucci C, Moltrasio M, et al. Acute kidney injury and in-hospital mortality in patients with ST-elevation myocardial infarction of different age groups. Int J Cardiol. Dec 01, 2021;344:8-12. [CrossRef] [Medline]
    38. Căruntu F, Bordejevic DA, Buz B, Gheorghiu A, Tomescu MC. Independent predictors of in-hospital and 1-year mortality rates in octogenarians with acute myocardial infarction. Rev Cardiovasc Med. Jun 30, 2021;22(2):489-497. [CrossRef] [Medline]
    39. Lv J, Wang C, Gao X, Yang J, Zhang X, Ye Y, et al. China Acute Myocardial Infarction Registry study group. Development and validation of dynamic models to predict postdischarge mortality risk in patients with acute myocardial infarction: results from China Acute Myocardial Infarction Registry. BMJ Open. Mar 29, 2023;13(3):e069505. [CrossRef] [Medline]
    40. Li L, Zhang X, Wang Y, Yu X, Jia H, Hou J, et al. A novel risk score to predict in-hospital mortality in patients with acute myocardial infarction: results from a prospective observational cohort. Front Cardiovasc Med. 2022;9:840485. [CrossRef] [Medline]
    41. Tan L, Xu Q, Shi R. A nomogram for predicting hospital mortality in intensive care unit patients with acute myocardial infarction. Int J Gen Med. 2021;14:5863-5877. [CrossRef] [Medline]
    42. Mizuguchi Y, Konishi T, Nagai T, Sato T, Takenaka S, Tada A, et al. Prognostic value of admission serum magnesium in acute myocardial infarction complicated by malignant ventricular arrhythmias. Am J Emerg Med. Jun 2021;44:100-105. [CrossRef] [Medline]
    43. Ravi D, Wong C, Deligianni F, Berthelot M, Andreu-Perez J, Lo B, et al. Deep learning for health informatics. IEEE J Biomed Health Inform. Jan 2017;21(1):4-21. [CrossRef] [Medline]
    44. Egger J, Gsaxner C, Pepe A, Pomykala KL, Jonske F, Kurz M, et al. Medical deep learning-a systematic meta-review. Comput Methods Programs Biomed. Jun 2022;221:106874. [CrossRef] [Medline]
    45. Mohammad MA, Olesen KK, Koul S, Gale CP, Rylance R, Jernberg T, et al. Development and validation of an artificial neural network algorithm to predict mortality and admission to hospital for heart failure after myocardial infarction: a nationwide population-based study. Lancet Digit Health. Jan 2022;4(1):e37-e45. [CrossRef]
    46. Johnson A, Bulgarelli L, Pollard T, Horng S, Celi LA, Mark R. MIMIC-IV (version 2.2). PhysioNet. Jan 6, 2023. URL: https://physionet.org/content/mimiciv/2.2/ [accessed 2024-04-29]
    47. eICU-CRD v2.0 documentation. eICU Collaborative Research Database. URL: https://eicu-crd.mit.edu/about/eicu/ [accessed 2024-04-29]

    ‎
    AMI: acute myocardial infarction
    AUC: area under the receiver operating characteristic curve
    eICU-CRD: Electronic Intensive Care Unit Collaborative Research Database
    ICD: International Classification of Diseases
    ICD-9: International Classification of Diseases, Ninth Revision
    ICD-10: International Statistical Classification of Diseases, Tenth Revision
    KNN: K-nearest neighbors
    LR: logistic regression
    MIMIC-IV: Medical Information Mart for Intensive Care-IV
    MLP: multilayer perceptron
    RF: random forest
    SAINT: Self-Attention and Intersample Attention Transformer
    SHAP: Shapley Additive Explanations
    XGBoost: extreme gradient boosting

    Edited by T de Azevedo Cardoso; submitted 12.06.23; peer-reviewed by S Schallmoser, JW Li, Y Wei; comments to author 14.09.23; revised version received 02.10.23; accepted 02.04.24; published 10.05.24.

    Copyright

    ©Puguang Xie, Hao Wang, Jun Xiao, Fan Xu, Jingyang Liu, Zihang Chen, Weijie Zhao, Siyu Hou, Dongdong Wu, Yu Ma, Jingjing Xiao. Originally published in the Journal of Medical Internet Research (https://www.jmir.org), 10.05.2024.

    This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in the Journal of Medical Internet Research, is properly cited. The complete bibliographic information, a link to the original publication on https://www.jmir.org/, as well as this copyright and license information must be included.