Skip to main content
Browse Subject Areas

Click through the PLOS taxonomy to find articles in your field.

For more information about PLOS Subject Areas, click here.

  • Loading metrics

Deep-learning-based risk stratification for mortality of patients with acute myocardial infarction

  • Joon-myoung Kwon ,

    Contributed equally to this work with: Joon-myoung Kwon, Ki-Hyun Jeon

    Roles Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Project administration, Resources, Software, Validation, Visualization, Writing – original draft, Writing – review & editing

    Affiliations Department of Emergency Medicine, Mediplex Sejong Hospital, Incheon, Korea, Artificial intelligence and big data center, Sejong medical research institute, Gyeonggi, Korea

  • Ki-Hyun Jeon ,

    Contributed equally to this work with: Joon-myoung Kwon, Ki-Hyun Jeon

    Roles Conceptualization, Data curation, Formal analysis, Funding acquisition, Methodology, Project administration, Supervision, Validation, Visualization, Writing – original draft, Writing – review & editing

    Affiliations Artificial intelligence and big data center, Sejong medical research institute, Gyeonggi, Korea, Division of Cardiology, Cardiovascular Center, Mediplex Sejong Hospital, Incheon, Korea

  • Hyue Mee Kim,

    Roles Conceptualization, Data curation, Methodology, Resources, Writing – review & editing

    Affiliation Division of Cardiology, Cardiovascular Center, Mediplex Sejong Hospital, Incheon, Korea

  • Min Jeong Kim,

    Roles Conceptualization, Data curation, Formal analysis, Methodology, Validation, Writing – review & editing

    Affiliation Division of Cardiology, Cardiovascular Center, Mediplex Sejong Hospital, Incheon, Korea

  • Sungmin Lim,

    Roles Data curation, Investigation, Writing – review & editing

    Affiliation Division of Cardiology, Cardiovascular Center, Mediplex Sejong Hospital, Incheon, Korea

  • Kyung-Hee Kim,

    Roles Data curation, Investigation, Methodology, Resources

    Affiliations Artificial intelligence and big data center, Sejong medical research institute, Gyeonggi, Korea, Division of Cardiology, Cardiovascular Center, Mediplex Sejong Hospital, Incheon, Korea

  • Pil Sang Song,

    Roles Conceptualization, Data curation, Funding acquisition, Resources, Writing – review & editing

    Affiliation Division of Cardiology, Cardiovascular Center, Mediplex Sejong Hospital, Incheon, Korea

  • Jinsik Park,

    Roles Conceptualization, Funding acquisition, Resources, Supervision, Writing – review & editing

    Affiliation Division of Cardiology, Cardiovascular Center, Mediplex Sejong Hospital, Incheon, Korea

  • Rak Kyeong Choi,

    Roles Conceptualization, Methodology, Resources, Validation

    Affiliation Division of Cardiology, Cardiovascular Center, Mediplex Sejong Hospital, Incheon, Korea

  • Byung-Hee Oh

    Roles Conceptualization, Data curation, Methodology, Resources, Writing – review & editing

    Affiliation Division of Cardiology, Cardiovascular Center, Mediplex Sejong Hospital, Incheon, Korea



Conventional risk stratification models for mortality of acute myocardial infarction (AMI) have potential limitations. This study aimed to develop and validate deep-learning-based risk stratification for the mortality of patients with AMI (DAMI).


The data of 22,875 AMI patients from the Korean working group of the myocardial infarction (KorMI) registry were exclusively divided into 12,152 derivation data of 36 hospitals and 10,723 validation data of 23 hospitals. The predictor variables were the initial demographic and laboratory data. The endpoints were in-hospital mortality and 12-months mortality. We compared the DAMI performance with the global registry of acute coronary event (GRACE) score, acute coronary treatment and intervention outcomes network (ACTION) score, and the thrombolysis in myocardial infarction (TIMI) score using the validation data.


In-hospital mortality for the study subjects was 4.4% and 6-month mortality after survival upon discharge was 2.2%. The areas under the receiver operating characteristic curves (AUCs) of the DAMI were 0.905 [95% confidence interval 0.902–0.909] and 0.870 [0.865–0.876] for the ST elevation myocardial infarction (STEMI) and non ST elevation myocardial infarction (NSTEMI) patients, respectively; these results significantly outperformed those of the GRACE (0.851 [0.846–0.856], 0.810 [0.803–0.819]), ACTION (0.852 [0.847–0.857], 0.806 [0.799–0.814] and TIMI score (0.781 [0.775–0.787], 0.593[0.585–0.603]). DAMI predicted 30.9% of patients more accurately than the GRACE score. As secondary outcome, during the 6-month follow-up, the high risk group, defined by the DAMI, has a significantly higher mortality rate than the low risk group (17.1% vs. 0.5%, p < 0.001).


The DAMI predicted in-hospital mortality and 12-month mortality of AMI patients more accurately than the existing risk scores and other machine-learning methods.


In the past decades, the mortality rate of acute myocardial infarction (AMI) has improved with advances in early reperfusion therapy and adjunctive pharmacotherapy.[1] However, AMI is still the major leading cause of mortality worldwide.[24] Risk stratification and prognosis prediction are critical in identifying high risk patients and decision making for the treatment of patients with AMI.[5] Conventional risk scoring systems including the thrombolysis in myocardial infarction (TIMI), the global registry of acute coronary events (GRACE), and the acute coronary treatment and intervention outcomes network (ACTION) scores are widely validated and accepted tools that are estimated using patients’ clinical information.[68] However, these prognostic models have limitations for the current daily practice. First, these systems are questionable in contemporary practice because they had been developed 20 years ago. Additionally, as these models use only selective variables based on a conventional statistical method, there is a possibility of loss of important information.[912]

Recently, deep-learning has achieved high performance in several medical domains, such as image classification (e.g., detection of abnormalities in retinal funduscopic result) and clinical outcome prediction (e.g., in-hospital mortality and long-term outcomes).[1315] An advantage of deep-learning is the automatic learning of the feature and relationship from a given data.[16] In this study, we developed and validated a deep-learning-based risk stratification for the mortality of patients with acute myocardial infarction (DAMI) using the Korean working group of myocardial infarction (KorMI) registry, a large national data.


Study population

We conducted a retrospective observational cohort study using data from the KorMI registry. KorMI is a prospective multicenter registry of AMI in Korean patients. All 59 cardiovascular centers in Korea were included in this study in January 2008. The full details of the KorMI registry’s aims and protocols have been published elsewhere.[17] This study was conducted in accordance with the Declaration of Helsinki and the relevant guidelines and regulations. The institutional review boards of Sejong General Hospital and Mediplex Sejong Hospital approved this study protocol and granted waivers of informed consent based on general impracticability and minimal harm. Patient information was anonymized and de-identified before the analysis. The data obtained through KorMI were the demographic information, treatment in the emergency room, laboratory results, electrocardiography findings, final diagnosis, clinical outcome during their hospital stay, and 12-month prognosis after discharge. The data were collected at each hospital using an encrypted web database.

Data management

First, the study data was split according to the hospital to prepare the validation data (Fig 1). The number of hospitals contributing data to the KorMI registry was 59, and we randomly selected 60% (36 hospitals) and 40% (23 hospitals) of these hospitals from which to obtain training data and validation data, respectively. As deep-learning and machine-learning can learn the characteristics of a hospital, we divided the data into training and validation dataset not by study subjects, but by hospitals, for the purpose of confirming further exact validation tests. For more accurate validation tests and subgroup analyses, we distributed as many study subjects as possible into the validation datasets. The validation data was used to confirm whether the DAMI can be applied to other hospitals after development.

The DAMI is a risk stratification model to predict in-hospital mortality after an AMI. We used the demographic information and laboratory data of AMI patients including age, sex, body mass index, cardiac arrest before visit, systolic blood pressure, heart rate, Killip class, creatinine kinase-muscle/brain (CK-MB), glucose, C-reactive protein (CRP), creatinine, low-density lipoprotein, and elevation of the ST segment, as the predictor variables. We aim to develop the model to help physicians in deciding a treatment plan, such as performing emergent coronary angiography or mechanical circulatory support, at the time of initial evaluation and treatment. Because of this, we used the predictor variables which could be obtained at the time of initial evaluation. We used first documented values in each admission, such as a first vital heart rate at emergency department.

Development of machine-learning prediction model

As shown in Fig 2, we developed the DAMI using only the derivation data. The The DAMI is a multilayer perceptron (MLP) based on deep-learning, and the DAMI incorporates four hidden layers, 102 nodes, batch normalization, and dropout layers.[1820] Because the accuracy did not increase when five or more hidden layers were added, we used four hidden layers to minimize the number of parameters to be learned. We used Tensor Flow (the Google Brain Team) as the backend.[21] Further, we used the Adagrad optimizer with the default parameters and binary-cross entropy as the loss function.[22] One node of the MLP is added by multiplying the values from the upper layer nodes (xk) by the weights (wk). The added value, (x1w1+x2w2+⋯+xkwk+c), is processed by the activation function, and the value f(x1w1+x2w2+⋯+xkwk+c) is sent to the next node. In this MLP, we used a rectified linear unit (ReLU) as the activation function.[23] Because maximum accuracy is observed for this predictive model using ReLU when compared with other activation functions such as soft max, linear, Tanh, leaky ReLU, and exponential linear unit, all the hyper-parameters used in the DAMI were tuned using grid search and manual tuning.

Fig 2. Deep-learning based model development and accuracy test.

AMI denotes acute myocardial infarction, CKMB creatinine kinase-MB, CVA cerebrovascular accident, DM diabetes mellitus, HTN hypertension, PMHx past medical history, ReLU rectified linear unit.

Before using the derivation data for the model development, we replaced the values of the categorical variables to numeric values and normalized the values of the continuous variables (Fig 2).[24] This data preprocessing was performed in the derivation data and validation data, separately. To train the model, we input each value of the derivation data in the input layer and adjusted the weight (wk) using the back propagation.[25] We have provided our prediction model as S1 File.

We also develop two machine-learning models: logistic regression (LR) and random forest (RF), for the performance comparison with the DAMI.[26] In the previous studies, LR and RF are the most typically used machine-learning methods and showed better performance than traditional methods in several medical domains.[27,28] The RF model consisted of 10,000 decision trees using the “randomForest” package in R (R Development Core Team, Vienna, Austria).[29,30] The LR model was derived using the “‘glmulti” packages in R.[31] We used the original Akaike IC as the information criterion and forward-backward directions for LR model selection.

Validation of prediction model performance

After we developed the DAMI, LR, and RF models, we compared the performance of these models with the GRACE, ACTION, and TIMI scores. We compared the performance of the models using only validation data that were not used for the model development (Fig 1). We analyzed the variable importance of logistic regression, random forest, and deep-learning by using deviance difference, mean decrease Gini, and AUC difference, respectively. In the GRACE and TIMI score, the formulas for calculating the risk score differ depending on the elevation of ST segment. For this reason, we divided the validation data into ST elevation myocardial infarction (STEMI) and non-ST elevation myocardial infarction (NSTEMI) and confirmed the accuracy at each group. We used the area under the receiver operating characteristic curve (AUC) as the comparative measure. [32]

We divided the patients of the validation data into high risk, intermediate risk, and low risk groups according to the DAMI and GRACE scores. The cutoff points of GRACE score were determined in previous studies.[33] The predicted mortality of low, intermediate, high risk group of GRACE score are less than 1%, 1–3%, and over 3%, respectively. And the optimal cutoff points of DAMI score were determined when the predicted in-hospital mortality of each risk group was equal to the that of GRACE score. After dividing the risk group by the DAMI and GRACE scores, we compared the accuracy for the in-hospital mortality through the reclassification table. Further, we confirmed characteristics of the DAMI risk groups. The continuous variables were presented as the mean and standard deviation and were compared using the analysis of variance (ANOVA) test. The categorical variables are expressed as frequencies and percentages and were compared by the Chi-square statistics. We estimated the 6-month mortality rate by the DAMI risk groups using the Kaplan–Meier method.


We included 25,977 AMI patients enrolled in the KorMI registry from January 2008 to December 2013 and excluded 3,102 patients according to the criteria mentioned in Fig 1. There is observed to be no significant difference in predictor variables between included and excluded study subjects, as shown in S1 Table. The study subjects comprised 22,875 patients of 59 hospitals, where 1,081 had in-hospital mortality. In-hospital mortality for the study subjects was 4.4%. In study subjects who survived upon discharge, 6-month mortality was 2.2% and mean (± standard deviation) of time after discharge was 28.5 days (± 41.7 days). The DAMI was developed using 12,152 patients of the derivation data. The accuracy test was performed using 10,723 patients of the validation data, where STEMI and NSTEMI patients were 5,841 and 4,882, respectively (Fig 1). We provide the developed DAMI, coding book for making input tidy data, example of tidy validation data, and python code for accuracy test as a S1 File to this article.

As shown in Fig 3, during the accuracy test of STEMI patients, the AUC of the DAMI was 0.905 [95% confidence interval 0.902–0.909] and this result significantly outperformed the GRACE score (0.851 [0.846–0.856]), ACTION score (0.852 [0.847–0.857]), TIMI score (0.781 [0.775–0.787]), LR (0.873 [0.869–0.878]), and RF (0.890 [0.886–0.895]). In the NSTEMI patients group, the AUC of the DAMI was 0.870 [0.865–0.876] and this accuracy significantly outperformed the GRACE score (0.810 [0.803–0.819]), ACTION score (0.806 [0.799–0.814]), TIMI score (0.593 [0.585–0.603]), LR (0.845 [0.839–0.851]), and RF (0.851 [0.845–0.858]). The variable importance of each prediction model is shown in S2 Table. The variable importance is different for each prediction model.

Fig 3. Receiver operating characteristic curve for predicting in-hospital mortality.

AUC denotes area under the receiver operating characteristic curve, CI confidence interval, GRACE global registry of acute coronary event, TIMI thrombolysis in myocardial infarction.

In the following experiments, we used the combined data from the STEMI and NSTEMI validation data. In the validation data, the cut-off scores of the DAMI risk groups were 2.3 and 7.9. With this cut-off value, the DAMI classified 2,843, 2,957, and 4,923 patients as high, intermediate, and low risk, respectively. Table 1 shows the baseline characteristics of patients between the high, intermediate, and low risk groups, defined by the DAMI. As shown in the reclassification table (Table 2), the DAMI predicted 34 in-hospital mortality patients and 3678 survival discharge patients more accurately than the GRACE score. And the DAMI predicted 31 in-hospital mortality patients and 573 survival discharge patients more incorrectly than the GRACE score. In 3526 patients who were in intermediate group by GRACE score, 1937 patients were reclassified to low risk group and in 50 patients of in-hospital death, 24 patients were reclassified to high risk group and 9 patients were reclassified to low risk group by DAMI score (Fig 4).

Fig 4. Reclassification of Individuals predicted to be at intermediate risk group by additional assessment of DAMI.

DAMI denotes deep-learning-based risk stratification for the mortality of patients with AMI and GRACE denotes global registry of acute coronary event.

For the analysis of mortality during the 6-month period, we considered the validation data of 10,723 patients (Fig 1). As shown in the Kaplan-Meier survival curves of Fig 5, the high-risk group defined by the DAMI shows a significantly higher hazard than the low risk group. The high-risk group, defined by the DAMI, has a significantly higher mortality rate than the low-risk group (17.1% vs. 0.5%, p < 0.001).

Fig 5. Kaplan-Meier survival curve stratified by deep-learning model risk score group.


In this study, we developed a risk stratification model for the mortality of patients with AMI using deep-learning from a large prospective national registry. By the accuracy test, this study revealed that the accuracy performance of the deep-learning model was excellent for predicting the prognosis and is better than the conventional risk-prediction model. To the best of our knowledge, this study is the first to predict AMI patient outcomes using deep-learning.

The TIMI and GRACE scores are extensively validated, traditional models for risk stratification following AMI.[33,34] The previous validation studies have reported that the AUCs of the TIMI and GRACE scores were 0.60–0.70 and 0.80–0.85, respectively.[35] It is confirmed in this large study population as 0.59–0.78 for the TIMI score and 0.81–0.85 for the GRACE score, implying moderate accuracy for predicting the mortality of AMI patients.

However, several notable limitations exist in the TIMI and GRACE scores. First, these models were developed based on the AMI patients’ data between the 1990s and early 2000s. In the past decade, early reperfusion therapy and the routine usage of drug-eluting stents have become routine. The benefit of intensive statin therapy was confirmed and potent antiplatelet agents, such as prasugrel and ticagrelor were introduced in our daily practice.[3639] In addition, these scoring systems are different for STEMI and NSTEMI. Recently, Song et al. reported a new scoring system for predicting the outcomes in survivors treated with guideline-adherent optimal therapies after AMI using the conventional statistical approach, which had better discrimination power than the GRACE model or other scoring system.[40] The DAMI risk stratification model was developed based on relatively recent data, which can better predict the mortality of AMI patients in the current practice. And the status of ST segment is included in DAMI algorithm, DAMI can equally well predict the mortality of AMI regardless of ST elevation. Next, the old models, used in TIMI, GRACE, and ACTION, inevitably restrict the numbers of predictive factors, because these models were developed by the conventional statistical approach using the logistic regression model that contains limitation including the fixed assumptions on data behavior, and the necessity to preselect variables in the development phase, thus leading to potential information loss.[912] Unlike the conventional statistical approach, deep-learning does not require the preselection of important variables, and the less important variables are naturally ignored in the model fitting.[4143] Further, deep-learning does not limit the number of input predictive factors and can use all available information without potential loss. Subsequently, the old models cannot reflect the relationship between variables. This is because the risk is measured only by the sum of the variables. Meanwhile, deep-learning obtains the relationship between the variables, as shown in Fig 2, unlike conventional methods.[16]

A previous study attempted to predict a 30-day mortality after ST-elevation myocardial infarction using conventional machine-learning methods including LR and RF and confirmed that RF performed the best.[28] However, no significant difference in performance is shown between the RF and GRACE scores. In that study, the machine-learning requires a feature-selection step before developing a predictive model. The feature selection is to delete variables that are less relevant to the prediction outcome and leads to potential information loss.[44] An important advantage of the deep-learning compared with conventional machine-learning, such as LR and RF, is feature learning.[16] In our study, feature learning is applied to obtain useful features to predict the endpoint of an AMI patient. Using a large amount of data, the deep-learning model automatically learns the features and conducts the given tasks such as classification and detection. This is why deep-learning shows better results than traditional machine-learning.[4143]

Deep-learning and machine-learning are used to obtain the relationship between the predictor variables and outcome variable, rather than creating a rule based on medical knowledge. Hence, the performance of machine-learning and deep-learning is not guaranteed in other situations as the algorithms can memorize only the characteristics of the derivation data. Wolpert explains the no-free-lunch theorem; if optimized in one situation, a model cannot produce good results in other situations.[45] Hence, we conducted an accuracy test using data which were not used for the model derivation. As deep-learning and machine-learning can learn the characteristics of a hospital, the hospital that developed the model and the hospital that conducted the accuracy test were completely separated.

Many researchers have attempted to determine whether machine-learning models developed for the prediction of one outcome can predict other similar outcomes. For example, some researchers have confirmed that a machine-learning model trained from in-hospital cardiac arrest data can predict unexpected intensive care unit transfer due to deterioration or death without attempted resuscitation.[15,27] We have confirmed that DAMI, which was developed with in-hospital mortality data, can predict 6-month mortality in this study. Because the available data is limited and the outcomes to be predicted are highly diverse, this result is promising to future studies in medical domains and will inspire many researchers.

Several limitations are present in our study. First, deep-learning is known as a “black box.” Although we can fit the deep-learning model by confirming each weight (wk), we cannot interpret the deep-learning model, in terms of variable importance or the approach to the decision of risk score. Recently, interpretable deep-learning has been studied and will be our next area of study.[46,47] Second, as previously described, deep-learning models rely on the representability of data. One of the most important characteristics of deep-learning is that it uses only the relationship between variables, as opposed to medical knowledge. Because of this, the developed deep-learning-based model can be tied to representativeness of training data and can thus be biased. Hence, it is necessary to validate this model in other environments, we have provided our prediction model as S1 File. Third, deep-learning only uses existing relationships in the data, regardless of whether they are due to causality or not. Fourth, as DAMI cannot be calculated manually, it is more difficult to use than conventional methods such as TIMI and GRACE. However, there are many hospitals using electronic health records (EHR), the DAMI score could be implemented to these EHRs with the S1 File and calculated automatically. Finally, the proportion of STEMI patients in this study was seen to be significantly higher than that of other studies, in which STEMI patients were 30–40%. There could be a risk of bias and overfitting for STEMI. However, the analysis of each group (STEMI and NSTEMI) showed similar patterns of the performance of deep learning model and conventional model for STEMI and NSTEMI.


In conclusion, we developed and validated a new risk stratification model of AMI based on the deep-learning approach. The DAMI predicted the in-hospital mortality and 12-month mortality of AMI patients more accurately than the existing risk scores and other machine-learning methods. This study showed the feasibility and effectiveness of the deep-learning-based algorithm model for cardiology, which can be a useful tool for precise decision making in daily practice.

Supporting information

S1 Table. Difference in predictor variables between included and excluded study subjects.


S2 Table. Importance of variables in derivation data for each algorithm.

AUC denotes area under the receiver operating characteristic curve, BMI body mass index, CRP C-reactive protein, CK-MB creatinine kinase-muscle/brain, LDL low density lipoprotein, OHCA out-of-hospital cardiac arrest, and SBP systolic blood pressure.


S1 File. Deep-learning-based risk stratification for the mortality of patients with acute myocardial infarction and materials for validation process.



  1. 1. Rahimi K, Duncan M, Pitcher A, Emdin CA, Goldacre MJ. Mortality from heart failure, acute myocardial infarction and other ischaemic heart disease in England and Oxford: a trend study of multiple-cause-coded death certification. J Epidemiol Community Health. 2015;69: 1000–5. pmid:26136081
  2. 2. Moran AE, Tzong KY, Forouzanfar MH, Rothy GA, Mensah GA, Ezzati M, et al. Variations in ischemic heart disease burden by age, country, and income: the Global Burden of Diseases, Injuries, and Risk Factors 2010 study. Glob Heart. 2014;9: 91–9. pmid:24977114
  3. 3. Pilgrim T, Vranckx P, Valgimigli M, Stefanini GG, Piccolo R, Rat J, et al. Risk and timing of recurrent ischemic events among patients with stable ischemic heart disease, non-ST-segment elevation acute coronary syndrome, and ST-segment elevation myocardial infarction. Am Heart J. 2016;175: 56–65. pmid:27179724
  4. 4. Arnold S V, Smolderen KG, Kennedy KF, Li Y, Shore S, Stolker JM, et al. Risk factors for rehospitalization for acute coronary syndromes and unplanned revascularization following acute myocardial infarction. J Am Heart Assoc. 2015;4. pmid:25666368
  5. 5. Yanishi K, Nakamura T, Nakanishi N, Yokota I, Zen K, Yamano T, et al. A Simple Risk Stratification Model for ST-Elevation Myocardial Infarction (STEMI) from the Combination of Blood Examination Variables: Acute Myocardial Infarction-Kyoto Multi-Center Risk Study Group. PLoS One. 2016;11: e0166391. pmid:27835698
  6. 6. Antman EM, Cohen M, Bernink PJ, McCabe CH, Horacek T, Papuchis G, et al. The TIMI risk score for unstable angina/non-ST elevation MI: A method for prognostication and therapeutic decision making. JAMA. 2000;284: 835–42. Available: pmid:10938172
  7. 7. Eagle KA, Lim MJ, Dabbous OH, Pieper KS, Goldberg RJ, Van de Werf F, et al. A validated prediction model for all forms of acute coronary syndrome: estimating the risk of 6-month postdischarge death in an international registry. JAMA. 2004;291: 2727–33. pmid:15187054
  8. 8. McNamara RL, Kennedy KF, Cohen DJ, Diercks DB, Moscucci M, Ramee S, et al. Predicting In-Hospital Mortality in Patients With Acute Myocardial Infarction. J Am Coll Cardiol. 2016;68: 626–635. pmid:27491907
  9. 9. Sun G, Shook TL, Kay GL. Inappropriate Use of Bivariable Analysis to Screen Risk Factors for Use in Multivariable Analysis. 1996;49: 907–916. pmid:8699212
  10. 10. Breiman L. Statistical Modeling: The Two Cultures (with comments and a rejoinder by the author). Stat Sci. 2001;16: 199–231.
  11. 11. Bagley SC, White H, Golomb BA. Logistic regression in the medical literature: standards for use and reporting, with particular attention to one medical domain. J Clin Epidemiol. 2001;54: 979–85. Available: pmid:11576808
  12. 12. Mining D. Data Mining: Statistics and More? 1998;52.
  13. 13. Gulshan V, Peng L, Coram M, Stumpe MC, Wu D, Narayanaswamy A, et al. Development and Validation of a Deep Learning Algorithm for Detection of Diabetic Retinopathy in Retinal Fundus Photographs. Jama. 2016;304: 649–656. pmid:27898976
  14. 14. Esteva A, Kuprel B, Novoa RA, Ko J, Swetter SM, Blau HM, et al. Dermatologist-level classification of skin cancer with deep neural networks. Nature. Nature Publishing Group; 2017;542: 115–118. pmid:28117445
  15. 15. Kwon J-M, Lee Y, Lee Y, Lee S, Park J. An Algorithm Based on Deep Learning for Predicting In-Hospital Cardiac Arrest. J Am Heart Assoc. 2018;7: e008678. pmid:29945914
  16. 16. LeCun Y, Bengio Y, Hinton G. Deep learning. Nature. 2015;521: 436–444. pmid:26017442
  17. 17. Sim DS, Jeong MH, Kang JC. Current management of acute myocardial infarction: Experience from the Korea Acute Myocardial. J Cardiol. Japanese College of Cardiology; 2010;56: 1–7. pmid:20554156
  18. 18. Schalkoff RJ. Pattern recognition—statistical, structural and neural approaches. [Internet]. New York: Wiley; 1992. Available:
  19. 19. Ioffe S, Szegedy C. Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift. 2015;
  20. 20. Srivastava N, Hinton G, Krizhevsky A, Sutskever I, Salakhutdinov R. Dropout: A Simple Way to Prevent Neural Networks from Overfitting. J Mach Learn Res. 2014;15: 1929–1958.
  21. 21. Abadi M, Barham P, Chen J, Chen Z, Davis A, Dean J, et al. TensorFlow: A System for Large-Scale Machine Learning TensorFlow: A system for large-scale machine learning. 12th USENIX Symp Oper Syst Des Implement (OSDI ‘16). 2016; 265–284.
  22. 22. Kingma DP, Ba J. Adam: A Method for Stochastic Optimization. 2017 IEEE Int Conf Consum Electron ICCE 2017. 2014; 434–435.
  23. 23. Nair V, Hinton GE. Rectified Linear Units Improve Restricted Boltzmann Machines. Proc 27th Int Conf Mach Learn. 2010; 807–814.
  24. 24. Jayalakshmi T, Santhakumaran A. Statistical Normalization and Backpropagation for Classification. Int J Comput Theory Eng. 2011;3: 89–93.
  25. 25. RUMELHART DE, HINTON GE, WILLIAMS RJ. Learning Internal Representations by Error Propagation. Readings in Cognitive Science. Elsevier; 1988. pp. 399–421.
  26. 26. Kuhn M, Johnson K. Applied Predictive Modeling [Internet]. New York, NY: Springer New York; 2013.
  27. 27. Churpek MM, Yuen TC, Winslow C, Meltzer DO, Kattan MW, Edelson DP. Multicenter Comparison of Machine Learning Methods and Conventional Regression for Predicting Clinical Deterioration on the Wards. Crit Care Med. 2016;44: 368–74. pmid:26771782
  28. 28. Shouval R, Hadanny A, Shlomo N, Iakobishvili Z, Unger R, Zahger D, et al. Machine learning for prediction of 30-day mortality after ST elevation myocardial infraction: An Acute Coronary Syndrome Israeli Survey data mining study. Int J Cardiol. Elsevier Ireland Ltd; 2017;246: 7–13. pmid:28867023
  29. 29. Liaw a, Wiener M. Classification and Regression by randomForest. R news. 2002;2: 18–22.
  30. 30. Khalilia M, Chakraborty S, Popescu M. Predicting disease risks from highly imbalanced data using random forest. BMC Med Inform Decis Mak. BioMed Central Ltd; 2011;11: 51. pmid:21801360
  31. 31. Calcagno V, Mazancourt C De. glmulti: An R Package for Easy Automated Model Selection with (Generalized) Linear Models. J Stat Softw. 2010;34: 1–29.
  32. 32. Fawcett T. An introduction to ROC analysis. Pattern Recognit Lett. 2006;27: 861–874.
  33. 33. Elbarouni B, Goodman SG, Yan RT, Welsh RC, Kornder JM, Deyoung JP, et al. Validation of the Global Registry of Acute Coronary Event (GRACE) risk score for in-hospital mortality in patients with acute coronary syndrome in Canada. Am Heart J. 2009;158: 392–9. pmid:19699862
  34. 34. Fox KAA, Fitzgerald G, Puymirat E, Huang W, Carruthers K, Simon T, et al. Should patients with acute coronary disease be stratified for management according to their risk? Derivation, external validation and outcomes using the updated GRACE risk score. BMJ Open. 2014;4: e004425. pmid:24561498
  35. 35. D’Ascenzo F, Biondi-Zoccai G, Moretti C, Bollati M, Omedè P, Sciuto F, et al. TIMI, GRACE and alternative risk scores in Acute Coronary Syndromes: a meta-analysis of 40 derivation studies on 216,552 patients and of 42 validation studies on 31,625 patients. Contemp Clin Trials. 2012;33: 507–14. pmid:22265976
  36. 36. Yan YL, Qiu B, Hu LJ, Jing XD, Liu YJ, Deng SB, et al. Efficacy and safety evaluation of intensive statin therapy in older patients with coronary heart disease: a systematic review and meta-analysis. Eur J Clin Pharmacol. 2013;69: 2001–9. pmid:23942733
  37. 37. Bohula EA, Giugliano RP, Cannon CP, Zhou J, Murphy SA, White JA, et al. Achievement of dual low-density lipoprotein cholesterol and high-sensitivity C-reactive protein targets more frequent with the addition of ezetimibe to simvastatin and associated with better outcomes in IMPROVE-IT. Circulation. 2015;132: 1224–33. pmid:26330412
  38. 38. Wallentin L, Becker RC, Budaj A, Cannon CP, Emanuelsson H, Held C, et al. Ticagrelor versus clopidogrel in patients with acute coronary syndromes. N Engl J Med. 2009;361: 1045–57. pmid:19717846
  39. 39. Roe MT, Armstrong PW, Fox KAA, White HD, Prabhakaran D, Goodman SG, et al. Prasugrel versus clopidogrel for acute coronary syndromes without revascularization. N Engl J Med. 2012;367: 1297–309. pmid:22920930
  40. 40. Song PS, Ryu DR, Kim MJ, Jeon KH, Choi RK, Park JS, et al. Risk Scoring System to Assess Outcomes in Patients Treated with Contemporary Guideline-Adherent Optimal Therapies after Acute Myocardial Infarction. Korean Circ J. 2018;48: 492–504. pmid:29856143
  41. 41. Bengio Y, Courville A, Vincent P. Representation learning: A review and new perspectives. IEEE Trans Pattern Anal Mach Intell. 2013;35: 1798–1828. pmid:23787338
  42. 42. Mo D. A survey on deep learning: One small step toward AI. Tech Report, Univ New Mex Dept Comput Sci. 2012; 1–16. Available:
  43. 43. Bengio Y, {LeCun} Y, Lecun Y. Scaling Learning Algorithms towards AI. Large Scale Kernel Mach. 2007; 321–360.
  44. 44. Hall MA, Holmes G. Benchmarking attribute selection techniques for discrete class data mining. IEEE Trans Knowl Data Eng. 2003;15: 1437–1447.
  45. 45. Wolpert DH. The Supervised Learning No-Free-Lunch Theorems. Proc 6th Online World Conf Soft Comput Ind Appl. 2001; 10–24. Available:
  46. 46. Chen X, Duan Y, Houthooft R, Schulman J, Sutskever I, Abbeel P. InfoGAN: Interpretable Representation Learning by Information Maximizing Generative Adversarial Nets. 2016;
  47. 47. Fong RC, Vedaldi A. Interpretable Explanations of Black Boxes by Meaningful Perturbation. Proc IEEE Int Conf Comput Vis. 2017;2017-Octob: 3449–3457.