IJCA-28578; No of Pages 8

International Journal of Cardiology xxx (xxxx) xxx

Contents lists available at ScienceDirect

International Journal of Cardiology

journal homepage: www.elsevier.com/locate/ijcard

Early detection of ST-segment elevated myocardial infarction by artificial

intelligence with 12-lead electrocardiogram
Yifan Zhao a, Jing Xiong a, Yang Hou b, Mengyun Zhu a, Yuyan Lu a, Yuanxi Xu a, Jiadela Teliewubai a,
Weijing Liu a, Xiao Xu b, Xin Li b, Zheng Liu a, Wenhui Peng a, Xianxian Zhao c, Yi Zhang a,⁎, Yawei Xu a,⁎
a
Department of Cardiology, Shanghai Tenth People's Hospital, Tongji University School of Medicine, Shanghai, China
b
Shanghai MobileVision Technology Co., Ltd, China
c
Department of Cardiovasology, Changhai Hospital, Second Military Medical University, Shanghai, China

a r t i c l e i n f o a b s t r a c t

Article history: Patient delay is a worldwide unsolved problem in ST-segment elevated myocardial infarction (STEMI). An accu-
Received 1 April 2020 rate warning system based on electrocardiogram (ECG) may be a solution for this problem, and artificial intelli-
Accepted 30 April 2020 gence (AI) may offer a path to improve its accuracy and efficiency. In the present study, an AI-based STEMI
Available online xxxx
autodiagnosis algorithm was developed using a dataset of 667 STEMI ECGs and 7571 control ECGs. The algorithm
for detecting STEMI proposed in the present study achieved an area under the receiver operating curve (AUC) of
Keywords:
ST segment elevated myocardial infarction
0.9954 (95% CI, 0.9885 to 1) with sensitivity (recall), specificity, accuracy, precision and F1 scores of 96.75%,
Artificial intelligence 99.20%, 99.01%, 90.86% and 0.9372 respectively, in the external evaluation. In a comparative test with cardiolo-
Patient delay gists, the algorithm had an AUC of 0.9740 (95% CI, 0.9419 to 1), and its sensitivity (recall), specificity, accuracy,
precision, and F1 score were 90%, 98% and 94%, 97.82% and 0.9375 respectively, while the medical doctors had
sensitivity (recall), specificity, accuracy, precision and F1 score of 71.73%, 89.33%, 80.53%, 87.05% and 0.8817 re-
spectively. This study developed an AI-based, cardiologist-level algorithm for identifying STEMI.
© 2020 Elsevier B.V. All rights reserved.

1. Introduction this was especially surprising given that the door-to-balloon (D2B)
time, was previously below 90 min in China. Although Chest Pain Cen-
Cardiovascular disease (CVD) is the leading cause of mortality ters (CPCs) are spread widely throughout China to reduce the D2B
worldwide, accounting for approximately 17.3 million of 54 million time, the problem of excessive patient delay remains unsolved. This
total deaths and 31.5% of global all-cause mortality [1]. Among cardio- problem is not limited to China. The recent ACC/AHA STEMI guidelines
vascular diseases, ST-elevated myocardial infarction (STEMI), is the [2] and the recent European guidelines for coronary revascularization
most severe and challenging condition, with an in-hospital mortality [5] have both noted long patient delays and suggested that awareness
of approximately 5–6% in the United States [2]. In recent years, both of this issue should be increased, because it represents the leading un-
the incidence and mortality of STEMI have decreased, but the situation solved problem in the treatment of STEMI patients.
is still far from satisfactory, especially in China. According to a recent It is well established that successful treatment is time-dependent in
Chinese CVD report [3], CVD-related deaths account for up to 45.0% STEMI patients. It was previously reported that a change in D2B from 15
and 42.6% of all-cause mortality in rural and urban areas, respectively. to 180 min would lead to an increase in in-hospital mortality of 3% to
Moreover, CVD-related deaths in China increased by 46% from 1990 to 8.5% and an increase in 6-month mortality from 10% to 20% [6]. Consid-
2013. Deaths from ischemic heart disease (IHD) contributed the most, ering that electrocardiography (ECG) plays a crucial role in the early de-
with a rate of increase of 90.9%. This increasing mortality was largely re- tection of STEMI, an ECG monitor equipped with a specific algorithm to
lated to a significant time delay in treatment for STEMI patients. In a na- automatically detect ST segment elevation could theoretically help in
tional survey of 3434 AMI patients in China (the China-PEACE study), Li prealerting STEMI and reducing patient delay. However, this warning
et al. [4] reported that the average patient delay, which was defined as system would need to rely largely on an automatic diagnostic system
the period from symptom onset to hospital arrival, was as high as 4 h; that achieves high sensitivity and specificity. In recent years, some suc-
cessful ECG monitors with similar designs have achieved positive results
for the detection of arrhythmia patients, such as patients with atrial fi-
⁎ Corresponding authors at: Department of Cardiology, Shanghai Tenth People's
Hospital, Tongji University School of Medicine, 301 Middle Yanchang Road, Shanghai
brillation (AF). In the Rehearse-AF study, Halcox et al. [7] reported
200072, China. that among 1001 participants at high risk for stroke, the AF screening
E-mail addresses: yizshcn@gmail.com (Y. Zhang), xuyawei@tongji.edu.cn (Y. Xu). rate increased significantly when an AliveECG monitor equipped with

https://doi.org/10.1016/j.ijcard.2020.04.089
0167-5273/© 2020 Elsevier B.V. All rights reserved.

Please cite this article as: Y. Zhao, J. Xiong, Y. Hou, et al., Early detection of ST-segment elevated myocardial infarction by artificial intelligence with
12-lead..., International Journal of Cardiology, https://doi.org/10.1016/j.ijcard.2020.04.089
2 Y. Zhao et al. / International Journal of Cardiology xxx (xxxx) xxx

ECG interpretation software was used twice a week. In addition, the 2.4. Preprocessing
VEST study [8] reported that among 1475 AMI patients with left ventric-
ular ejection fraction (LVEF) b 35%, the 90-day all-cause mortality rate The sample dimension was 12 × 5000 it included 12 leads with 5000
was significantly reduced (by 35%) for patients who wore a cardiac de- points per lead. The raw ECG data contained a large amount of noise and
fibrillation (WCD) device equipped with an ECG diagnostic algorithm always suffered from baseline drift (Fig. 3A and B). Hence, we
compared with patients in a control group. However, automatically de- preprocessed all the raw ECG data before conducting training. To elim-
tecting arrhythmia is substantially different from detecting STEMI. The inate noise with low power and the baseline drift phenomenon from the
difficulties of detecting STEMI from ECG signals hinge on three factors. raw ECG data, we chose the well-known finite impulse response (FIR)
First, while single‑lead ECG is adequate for detecting arrhythmia, for and bandpass filter with a 3 × 45 window that was originally designed
STEMI alerting, at least a 12‑lead ECG is necessary. Therefore, the algo- by Yong and Choo [16]. After preprocessing, the noise and baseline drift
rithm is assumed to be much more complicated. Second, the algorithm were removed (see Fig. 3C and D).
for detecting STEMI should have a relatively higher sensitivity compared
with nonfatal arrhythmia detection because it must be designed to warn 2.5. Data expanding
of a potentially fatal condition, which needs to be unmissable. Third, the
ECG signal for detecting arrhythmia is relatively prominent, whereas Firstly, we down sample the original data from 500 Hz to 128 Hz by
that for detecting STEMI is less distinctive; moreover, the relatively using one dimensional interpolation and uniform sampling, so the data
tiny change in the ST-segment could easily be obliterated by a filtering changed to 12 × 1280 (12 leads with 1280 points per lead). Then, after
process. Thus, the raw ECG data without any filtering or visualization leaving out 100 samples (50 STEMI samples and 50 other samples) for
processes would be required to develop the STEMI algorithm. Consider- competition, we split the left 8138 samples into a training set and an ex-
ing these three problems, we realized that to solve this unmet clinical ternal validation set at proportions of 7:3 (the training set contained
need, artificial intelligence—more specifically, a deep convolutional neu- 5697 ECGs and the external validation set contained 2441 ECGs). And
ral network—should be applied to build this diagnostic algorithm be- then, in order to produce more data and meanwhile to avoid affecting
cause there are millions of interactive raw ECG data that need to be the evaluation of the external validation set, we only expand the train-
interpreted. We selected only STEMI patients with coronary angiogra- ing set by shifting the start point and choosing continuous 640 points of
phy (CAG) as a gold standard for the definition of STEMI to increase each lead which contains 1280 points in total, so it can be expanded to
the sensitivity and specificity of the algorithm and conducted a retro- 640 samples for each sample by shifting start point 640 times. After data
spective investigation with the aim of building a diagnostic algorithm expanding, the expanded training set contains 276,480 Positive (STEMI)
for detecting STEMI based on raw 12‑lead ECG in consecutive STEMI pa- samples and 3,369,600 Negative (the control ECGs) samples.
tients and controls recruited from 2 cardiac centers in Shanghai, China.
2.6. Model
2. Materials and methods
The model we used is depicted in Fig. 3E. In the model, we used 1-D
2.1. Study design operator for all of the convolution, max pooling and average pooling
layers, and we used ReLu as all of the activation layers. There are 16
The schematic workflow of the present study is shown in Fig. 1. The Res-Net blocks in total. The number of filters of convolution layers in
present study included 4 steps. The first step was data collection and la- the first block was 64, and it was doubled when the block was 3,4,6
beling. We collected STEMI and control ECGs from 2 local databases at a and 3. The kernel size in the convolution layers was 3, but it was 7 in
hospital in Shanghai. The collected ECGs and CAGs were evaluated and the first convolution layer. And all of the kernel initializer we used he-
labeled by expert cardiologists (Fig. 1A). Next, we trained our deep normal. The last dense layer we used soft-max as the activate function.
learning algorithm using a labeled training set and tuned it using a la-
beled internal validation set (Fig. 1B). Furthermore, the algorithm 2.7. Training and developing
underwent an external validation for model evaluation (Fig. 1C). Finally,
after the algorithm achieved satisfactory performance in the internal After finalizing the architecture of our model, we trained it using
and external validation, a comparative test was performed (Fig. 1D). 3,646,080 ECGs, including 276,480 STEMI and 3,369,600 control ECGs.
We chose Res-Net as our model to better fit the ECG model and avoid
2.2. Data collection and labeling overfitting. Categorical cross entropy loss was used as the loss function,
and we used Adam optimization method with learning rate beta-1,
For algorithm development and validation, the training ECG data beta-2 and clip norm of 0.001, 0.9, 0.999 and 1. To avoid training bias,
were retrospectively collected from inpatients in the cardiology depart- which can be caused by the imbalance of our training and validation
ment of Shanghai Tenth People's Hospital and Changhai Hospital from data (276,480 positive samples and 3,369,600 negative samples), we
2008 to 2018. All the ECGs were formatted as raw data and there used an input batch of 80 samples (40 positives and 40 negatives) in
were no other specific data requirements. the expanded training set to the model each time. When the models'
According to real-world clinical practices, STEMI samples all arise performance was pretty good in the internal validation set, which
from changes to the ST segment of the ECGs. We collected 667 STEMI contained 1000 samples randomly choose from the expanded training
samples validated by coronary angiography. In contrast, the control set, we saved out the models. About 96 h was took for training on
samples contain various changes, including changes in ST episode, P GeForce GTX 1080Ti Processors with 2 × 9592 MB memory.
wave, QRS complex, T wave, etc. Therefore, we needed many more con-
trol samples to cover all these other diseases (7571 samples). More de- 2.8. Comparative test
tailed information about data collection are available on the online only
supplementary data. 100 ECG images containing 50 STEMI ECGs and 50 control ECGs
were randomly selected. We performed a comparative test among our
2.3. Algorithm construction AI-based algorithm, a current commercial ECG algorithm and 15 doc-
tors. The commercial ECG algorithm is under software copyright and
Constructing our deep learning algorithm involved 4 main tasks: is widely used in China. The 15 doctors included 5 medical interns, 5 in-
Preprocessing, Data Expanding, Res-Net based model development, ternal medicine residents and 5 experienced cardiologists. The medical
training and tuning. interns were students who had completed their theoretical studies of

Please cite this article as: Y. Zhao, J. Xiong, Y. Hou, et al., Early detection of ST-segment elevated myocardial infarction by artificial intelligence with
12-lead..., International Journal of Cardiology, https://doi.org/10.1016/j.ijcard.2020.04.089
 Y. Zhao et al. / International Journal of Cardiology xxx (xxxx) xxx 3

Fig. 1. Schematic diagram of the study workflow. (A). A total of 10,375 ECGs (1802 candidate STEMI ECGs and 8652 control ECGs) were initially selected for validation. Then, 3 experienced
cardiologists evaluated the ECGs and further validated them by CAG. After validation, 667 STEMI ECGs and 7571 ECGs were ultimately used in the study. (B). The AI algorithm development
and training process. (C). Internal validation of the AI algorithm. (D). A real-world comparative test was performed to validate the performance of the AI algorithm.

cardiology and ECGs. The internal medicine residents were doctors who 2.9. Statistical analysis
held a medical license but had not majored in cardiology. The cardiolo-
gists were doctors who had at least 5 years of experience working in a A comparison of baseline characteristics between the STEMI and
cardiology department. Each doctor was asked to make a diagnosis as control groups was conducted using Student's t-tests. A value of
to whether each patient's condition was STEMI or NOT STEMI based P b .05 was considered to indicate a statistically significant difference.
on their interpretations of that patient's ECG. The ROC curve and AUC were obtained using Python 3.5 with the

Please cite this article as: Y. Zhao, J. Xiong, Y. Hou, et al., Early detection of ST-segment elevated myocardial infarction by artificial intelligence with
12-lead..., International Journal of Cardiology, https://doi.org/10.1016/j.ijcard.2020.04.089
4 Y. Zhao et al. / International Journal of Cardiology xxx (xxxx) xxx

modules “Matplotlib 3.0.2” and “roc” following previous studies [17,18]. (Fig. 2A and B). The patient demographics and the CAG characteristics
The CI of the AUC was calculated using Python 3.5 according to the pre- were listed in Supplementary Table 1. In accordance with the real-
viously published formula [19]. world situation, patients suffering from STEMI also present diabetes
and hypertension at a much higher proportion, both of which were
3. Results high-risk factors for STEMI. Meanwhile, the left anterior descending cor-
onary artery (LAD) and right coronary artery (RCA) were the two major
3.1. Patient selection and baseline characteristics culprit vessels, and they accounted for N92% of all culprit vessels, which
also conformed to the real-world situation [10]. All these data indicated
In previous AI-based ECG diagnosis studies, the ECGs were mostly that our dataset was reliable and similar to real-world clinical practice.
annotated by ECG experts or cardiologists. However, STEMI ECG annota-
tion by ECG experts is not reliable because ST segment elevation can be 3.2. The performance of the AI algorithm
caused by multiple diseases [9]. Therefore, we viewed the coronary an-
giography of all the candidate STEMI patients labeled by cardiologists. STEMI is too severe a disease to allow misdiagnoses. Meanwhile,
This process helped to ensure that all the ECGs used for algorithm train- overdiagnosis of some low-risk disease as STEMI would cause large
ing were true STEMI ECGs, which greatly improved the training amounts of needlessly wasted medical resources. Hence, we built 102
accuracy. different deep learning models to obtain an algorithm with both high
A total of 667 STEMI samples and 7571 control samples were col- sensitivity and specificity (Supplemental Table 2). Through the internal
lected for AI-based algorithm training, external validation and the com- validation process, Model 1 was selected as the algorithm version used
parative test. Of these, 432 STEMI ECGs which were expanded to for further external validation because of the highest F1 score (0.9371).
276,480 samples and 5265 control ECGs which were expanded to The internal evaluation demonstrated a sensitivity, specificity, accuracy,
3,369,600 samples were prepared for algorithm training and tuning, precision and F1 scores of 97.22%, 99.11%, 98.96%, 89.94% and 0.9344 re-
1000 samples were randomly choose from the expanded training set, spectively (see Fig. 4A). Subsequently, a total of 185 STEMI ECGs and
185 STEM ECGs and 2256 control ECGs were used for external valida- 2256 control ECGs, which were hitherto unseen by the model, were
tion. 50 STEMI ECGs and 50 control ECGs for comparative testing used for the external validation. The results were shown in Fig. 4B.

Fig. 2. Data collection and labeling. (A). The selection process of STEMI ECGs. Of the total ECGs collected, 891 were excluded because they showed no significant ST segment elevation; 212
were excluded because no CAG was performed; and 32 were excluded because no definite culprit vessels could be identified. A total of 667 STEMI ECGs were finally enrolled. (B). The
selection and classification of control ECGs: 1081 of the 8652 control ECGs were excluded for poor image quality or no definite diagnosis. The total 7571 adopted ECGs included 3350
normal ECGs and 4221 abnormal ECGs associated with different diseases.

Please cite this article as: Y. Zhao, J. Xiong, Y. Hou, et al., Early detection of ST-segment elevated myocardial infarction by artificial intelligence with
12-lead..., International Journal of Cardiology, https://doi.org/10.1016/j.ijcard.2020.04.089
 Y. Zhao et al. / International Journal of Cardiology xxx (xxxx) xxx 5

Fig. 3. Algorithm preprocessing and development: (A) and (B) show representative ECGs with interference and baseline drift, respectively; (C) and (D) show representative ECGs after
preprocessing. (E) shows the architecture of DenseNet. The raw ECG data are preprocessed by FIR and bandpass filter, which are followed by convolution and max pooling. Next, a
dense block with 5 layers plus a transition layer containing 5 operations (batch normalization, ReLU, convolution, dropout and average pooling) are passed through 3 times, followed
by a dense block with 16 layers and another transition layer with 5 operations (batch normalization, ReLU, global pooling, flatten, and linear).

The model achieved an area under the receiver operating curve (AUC) of Fig. 4C) achieved sensitivity, specificity, accuracy, precision and F1
0.9954 (95% CI, 0.9885 to 1) with sensitivity, specificity, accuracy, preci- score of 32.0%, 90.0%, 61.0%, 76.19% and 0.4702 respectively. The perfor-
sion and F1 scores of 96.75%, 99.20%, 99.01%, 90.86% and 0.9372% re- mances of the 15 doctors (shown as dots of different colors) achieved a
spectively in the external evaluation. These data indicated that our sensitivity of 71.73% (42%–94%), a specificity of 89.33 (82%–100%), an
deep learning algorithm was sufficiently accurate to compete with ex- accuracy of 80.53% (71%–93%), a precision of 87.05% and an F1 score
perienced cardiologists. Therefore, we performed a comparative test of 0.7865. The performances of the current commercial algorithm and
to elucidate whether the algorithm was suitable for clinical practice. A the doctors showed a relatively low sensitivity but a high specificity be-
test suite consisting of 50 validated STEMI ECGs and 50 control ECGs cause some ST segment elevations were not typical; thus, they tended to
(all the ECGs were new to both the algorithm and the 15 doctors) was diagnose those types of ECGs as normal. In contrast, the AI algorithm can
randomly sequenced. The algorithm and the doctors (who had three identify both atypical and slightly elevated ST segments, indicating that
levels of expertise: intern, physician residence or experienced cardiolo- an ECG with an AI autodiagnostic system for STEMI shows promise and
gist), were independently required to judge whether an ECG repre- value as a clinical application.
sented STEMI or NOT STEMI.
The results are presented in Fig. 4C. The algorithm achieved an AUC 4. Discussion
of 0.974 (95% CI, 0.9419 to 1) with a sensitivity of 90.0%, a specificity of
98.0% and an accuracy of 94.0%, a precision of 97.8%, and a F1 score of In this study, we developed an autodiagnosis algorithm for
0.9375. The commercial algorithm (shown as a green pentacle in STEMI using 667 STEMI ECG data and demonstrated that the

Please cite this article as: Y. Zhao, J. Xiong, Y. Hou, et al., Early detection of ST-segment elevated myocardial infarction by artificial intelligence with
12-lead..., International Journal of Cardiology, https://doi.org/10.1016/j.ijcard.2020.04.089
6 Y. Zhao et al. / International Journal of Cardiology xxx (xxxx) xxx

Fig. 4. The performance of the algorithm in the validation and comparative tests. (A). The ROC curve of the algorithm in the internal validation. The orange line represents the algorithm.
(B). The ROC curve of the algorithm in the external validation. The orange line represents the algorithm. (C). The performance of the algorithm in the comparative test. The orange line
represents the algorithm, the red points represent 5 experienced cardiologists, the blue points represent 5 physicians in residence, and the green points represent 5 interns, the green
pentacle represents the current commercial algorithm. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

algorithm achieves sensitivity, specificity and accuracy, precision possible to create a model that achieves both high sensitivity and
and F1 scores of 96.75%, 99.20%, 99.01%, 90.86% and 0.9371 respec- high specificity using a 12‑lead ECG diagnostic system with only a
tively in the external evaluation. These results reveal that it is small amount of data.

Please cite this article as: Y. Zhao, J. Xiong, Y. Hou, et al., Early detection of ST-segment elevated myocardial infarction by artificial intelligence with
12-lead..., International Journal of Cardiology, https://doi.org/10.1016/j.ijcard.2020.04.089
 Y. Zhao et al. / International Journal of Cardiology xxx (xxxx) xxx 7

This system is an AI-based 12‑lead regular ECG diagnosis system, and prophylactic tool for individuals at high risk of cardiovascular dis-
it may provide valuable information for the future development of ECG eases (Supplementary Fig. 1).
automatic diagnostic techniques based on deep learning technologies, Another potential application of this AI-based ECG algorithm is the
especially for diseases that require multi‑lead ECG for diagnosis, such using by paramedics in ambulance and physicians from the referred
as the position-oriented diagnosis of premature beats, ischemic heart hospital and clinics. It was reported that the inappropriate activation
disease and ventricular hypertrophy. Sengupta et al. [11] proposed a of catheter lab for STEMI was around 20–25% [20], which was obviously
machine learning-based approach to identify myocardial relaxation ab- a great waste of medical resources. Application of this AI-based algo-
normalities from 12‑lead ECGs. However, the ECG signal was first trans- rithm at a global cloud-base, have the potential to impower a wide
formed to color images by a continuous wavelet transform (CWT) signal range of physicians to more accurately diagnose STEMI on ECG and re-
processing system that is still not widely available in clinical practice. duced the inappropriate activation of catheter lab. If so, this AI-based
Our algorithm was developed based on regular 12‑lead ECG data, STEMI algorithm may contribute greatly to the improvement of the cur-
which is the most common auxiliary examination method used for di- rent STEMI system around world. However, before that, this AI ECG al-
agnosing cardiovascular diseases. In terms of STEMI, the patient delay gorithm should be validated in various ethnics, because in the present
times are still unacceptably long worldwide, averaging 4 h in China study, this algorithm only was validated in Chinese.
and 2 h in the US [2,3]. Using a 12‑lead ECG with an autodiagnostic sys- In summary, an auto-diagnostic algorithm to identify STEMI patients
tem could help both patients and doctors become aware of STEMI in its from 12‑lead ECG data using an AI technique was developed and vali-
early phase, which could further reduce patient delay times. In addition, dated in this study.
STEMI is a severe condition that should not be misdiagnosed. Hence, an
autodiagnostic system for STEMI must have high sensitivity. The sensi- Acknowledgments
tivity of our algorithm reached over 90%—significantly higher than
that achieved by 15 medical doctors (71.73%). These results indicate We would like to thank Yixing Zheng, Weilun Meng, Jiamin Tang,
that our AI-based ECG autodiagnostic system could be regularly utilized Ximin Fan, Song Zhao, Yaru Wang and Fangfang Xu for their support
to guide clinical work in the future. during data collection and labeling.
AI has been applied to clinical medicine in recent years. Gulshan
et al. [12] developed a deep learning algorithm to detect diabetic reti- Funding
nopathy using 9963 training images. Esteva et al. [13] created an AI-
based classification of skin cancer using N120,000 skin images. Both This work was supported by grants from: the National Key
studies achieved competitive accuracy. Nevertheless, it is difficult to ob- Technology R&D Program during the Thirteenth Five-year Plan Pe-
tain such large amounts of labeled data in most circumstances. Thus, it is riod (No. 2017YFC0111800) the Shanghai Municipal Government
essential to develop novel methods that achieve good performance (15GWZK1002) the National Natural Science Foundation of China
when trained with relatively small sample datasets. In this study, to ad- (Nos. 81470394 and 81770226 to Yawei Xu, No. 81670377 to Yi
dress small and imbalanced samples (667 positive samples and 7571 Zhang, and No. 81700210 to Yifan Zhao).
negative samples were used to train the algorithm), we chose Res-Net
as our model and expanded the training set. Res-Net [14] has been Declaration of competing interest
widely used worldwide in medical fields, and it is useful for small sam-
ple cases. It has been demonstrated to be very efficient for numerous The authors declare that no competing financial interests exist.
medical problems, such as medical imaging segmentation [15]. How-
ever, the traditional Res-Net model is not suitable for diagnosing Appendix A. Supplementary data
STEMI based on 12‑lead ECG, and few related studies have previously
been published. Part of the reason for this lack of data may be that Supplementary data to this article can be found online at https://doi.
raw ECG data are single-dimension data, while Res-Net was originally org/10.1016/j.ijcard.2020.04.089.
designed to deal with general images with 2 or 3 dimensions. To fit
our data distribution, we choose a continuous 640 points from each References
lead, constructing 7680 total points and then reshaped the data to
[1] GBD 2017 Causes of Death Collaborators. Global, regional, and national age-sex-
12 × 640 as model input; consequently, the remaining 640 points
specific mortality for 282 causes of death in 195 countries and territories,
on each lead could be used to produce another 640 samples by mov- 1980–2017: A systematic analysis for the global burden of disease study 2017. Lan-
ing the start point of each lead. The real data acquired from the ECG cet. 2018; 392: 1736–1788.
collector leads were connected to different ECG starting points in the [2] P. T. O'Gara, F. G. Kushner, D. D. Ascheim, D. E. Casey, Jr., M. K. Chung, J. A. de Lemos,
S. M. Ettinger, J. C. Fang, F. M. Fesmire, B. A. Franklin, C. B. Granger, H. M. Krumholz, J.
patients. We hope that the novel methods used to establish this algo- A. Linderbaum, D. A. Morrow, L. K. Newby, J. P. Ornato, N. Ou, M. J. Radford, J. E.
rithm will accelerate the development of AI applications for ECG Tamis-Holland, C. L. Tommaso, C. M. Tracy, Y. J. Woo, D. X. Zhao, J. L. Anderson, A.
diagnosis. K. Jacobs, J. L. Halperin, N. M. Albert, R. G. Brindis, M. A. Creager, D. DeMets, R. A.
Guyton, J. S. Hochman, R. J. Kovacs, F. G. Kushner, E. M. Ohman, W. G. Stevenson,
In clinical practice, early STEMI detection is obviously of great clini- C. W. Yancy. 2013 ACCF/AHA guideline for the management of ST-elevation myocar-
cal significance because the very early stages of STEMI are the most vul- dial infarction: a report of the American college of cardiology foundation/American
nerable periods during which most sudden cardiac deaths occur [6]. heart association task force on practice guidelines. Circulation. 2018; 127: e36:2-
e425.
However, most patients suffering from ischemic symptoms do not call [3] C. Weiwei, G. Runlin, L. Lisheng, Z. Manlu, W. Wen, W. Yongjun, W. Zhaosu, L.
an emergency service immediately due to a lack of medical knowledge. Huijun, Z. Zhe, J. Lixin, H. Shengshou, Outline of the report on cardiovascular dis-
Our algorithm can provide more accurate information about whether eases in China, 2014, Eur. Heart J. Suppl. 18 (2016) F2–F11.
[4] J. Li, X. Li, Q. Wang, S. Hu, Y. Wang, F.A. Masoudi, J.A. Spertus, H.M. Krumholz, L. Jiang,
the patient is suffering from STEMI, which would be helpful in reducing
ST-segment elevation myocardial infarction in China from 2001 to 2011 (the China
the total ischemic time for STEMI patients. Meanwhile, it would also be PEACE-retrospective acute myocardial infarction study): a retrospective analysis of
useful for some high-risk patients with no or only slight ischemic symp- hospital data, Lancet. 385 (2015) 441–451.
[5] F.J. Neumann, M. Sousa-Uva, A. Ahlsson, F. Alfonso, A.P. Banning, U. Benedetto, R.A.
toms. For these reasons, we consider two groups of individuals as poten-
Byrne, J.P. Collet, V. Falk, S.J. Head, P. Juni, A. Kastrati, A. Koller, S.D. Kristensen, J.
tial candidates for our AI system. The first group includes patients Niebauer, D.J. Richter, P.M. Seferovic, D. Sibbing, G.G. Stefanini, S. Windecker, R.
experiencing acute myocardial infarction, with or without revasculari- Yadav, M.O. Zembala, ESO Scientific Document Group. 2018 ESC/EACTS guidelines
zation therapy, who are at high risk of cardiac reinfarction. The second on myocardial revascularization. Eur, Heart J. 40 (2019) 87–165.
[6] B. Ibanez, S. James, S. Agewall, M.J. Antunes, C. Bucciarelli-Ducci, H. Bueno, A.L.P.
group could be outfitted with a wearable ECG monitor equipped Caforio, F. Crea, J.A. Goudevenos, S. Halvorsen, G. Hindricks, A. Kastrati, M.J.
with an autodiagnosis system which could potentially be used as a Lenzen, E. Prescott, M. Roffi, M. Valgimigli, C. Varenhorst, P. Vranckx, P. Widimsky,

Please cite this article as: Y. Zhao, J. Xiong, Y. Hou, et al., Early detection of ST-segment elevated myocardial infarction by artificial intelligence with
12-lead..., International Journal of Cardiology, https://doi.org/10.1016/j.ijcard.2020.04.089
8 Y. Zhao et al. / International Journal of Cardiology xxx (xxxx) xxx

2017 ESC guidelines for the management of acute myocardial infarction in patients detection of diabetic retinopathy in retinal fundus photographs, JAMA 316 (2016)
presenting with ST-segment elevation: the task force for the management of acute 2402–2410.
myocardial infarction in patients presenting with ST-segment elevation of the [13] A. Esteva, B. Kuprel, R.A. Novoa, J. Ko, S.M. Swetter, H.M. Blau, S. Thrun,
European society of cardiology (ESC), Eur. Heart J. 39 (2018) 119–177. Dermatologist-level classification of skin cancer with deep neural networks, Nature
[7] J.P.J. Halcox, K. Wareham, A. Cardew, M. Gilmore, J.P. Barry, C. Phillips, M. Gravenor, 542 (2017) 115–118.
B., assessment of remote heart rhythm sampling using the AliveCor heart monitor to [14] He K, Zhang X, Ren S, et al. Deep Residual Learning for Image Recognition. 2016 IEEE
screen for atrial fibrillation: the REHEARSE-AF study, Circulation. 136 (2017) Conference on Computer Vision and Pattern Recognition (CVPR). (IEEE Computer Soci-
1784–1794. ety, 2016), pp. 770–778.
[8] J.E. Olgin, M.J. Pletcher, E. Vittinghoff, J. Wranicz, R. Malik, D.P. Morin, S. Zweibel, A.E. [15] H. Chen, Q. Dou, L. Yu, P. Heng, A. (2016). Voxresnet: deep voxelwise residual net-
Buxton, C.S. Elayi, E.H. Chung, E. Rashba, M. Borggrefe, T.F. Hue, C. Maguire, F. Lin, J.A. works for volumetric brain segmentation, NeuroImage 170 (2018) 446–455.
Simon, S. Hulley, B.K. Lee, Wearable cardioverter–defibrillator after myocardial in- [16] L. Yong, H. P. Choo. ECG noise reduction using multiplier-free FIR digital filters. Pro-
farction, N. Engl. J. Med. 379 (2018) 1205–1215. ceedings 7th International Conference on Signal Processing, 2004. (IEEE, Beijing, China,
[9] E.B. Hanna, D.L. Glancy, ST-segment elevation: Differential diagnosis, caveats, Cleve. 2004), pp. 2198–2201.
Clin. J. Med. 82 (2015) 373–384. [17] T. Fawcett. An introduction to ROC analysis. Pattern Recognit. Lett. 2016; 27:
[10] G.P. McCann, J.N. Khan, J.P. Greenwood, S. Nazir, M. Dalby, N. Curzen, S. 861–874.
Hetherington, D.J. Kelly, D.J. Blackman, A. Ring, C. Peebles, J. Wong, T. Sasikaran, [18] A.P. Bradley, The use of the area under the ROC curve in the evaluation of machine
M. Flather, H. Swanton, A.H. Gershlick, Complete versus lesion-only primary PCI: learning algorithms, Pattern Recogn. 30 (1997) 1145–1159.
the randomized cardiovascular MR CvLPRIT substudy, J. Am. Coll. Cardiol. 66 [19] C. Cortes, M. Mohri, Confidence intervals for the area under the ROC curve. Advances
(2015) 2713–2724. in Neural Information Processing Systems. (MIT Press, Cambridge, MA, 2005), pp.
[11] P.P. Sengupta, H. Kulkarni, J. Narula, Prediction of abnormal myocardial relaxation 305–312.
from signal processed surface ECG, J. Am. Coll. Cardiol. 71 (2018) 1650–1660. [20] I.C. Rokos, W.J. French, A. Mattu, G. Nichol, M.E. Farkouh, J. Reiffel, G.W. Stone, Ap-
[12] V. Gulshan, L. Peng, M. Coram, M.C. Stumpe, D. Wu, A. Narayanaswamy, S. propriate cardiac cath lab activation: optimizing electrocardiogram interpretation
Venugopalan, K. Widner, T. Madams, J. Cuadros, R. Kim, R. Raman, P.C. Nelson, J.L. and clinical decision-making for acute ST-elevation myocardial infarction,
Mega, D.R. Webster, Development and validation of a deep learning algorithm for American Heart Journa. 160 (2010) 995–1003.

Please cite this article as: Y. Zhao, J. Xiong, Y. Hou, et al., Early detection of ST-segment elevated myocardial infarction by artificial intelligence with
12-lead..., International Journal of Cardiology, https://doi.org/10.1016/j.ijcard.2020.04.089