Abstract
Background
Elevated troponin levels are a sensitive biomarker for cardiac injury. The quick and reliable prediction of troponin elevation for patients with chest pain from readily available ECGs may pose a valuable time-saving diagnostic tool during decision-making concerning this patient population.
Methods and results
The data used included 15 856 ECGs from patients presenting to the emergency rooms with chest pain or dyspnoea at two centres in Sweden from 2015 to June 2023. All patients had high-sensitivity troponin test results within 6 hours after 12-lead ECG. Both troponin I (TnI) and TnT were used, with biomarker-specific cut-offs and sex-specific cut-offs for TnI. On this dataset, a residual convolutional neural network (ResNet) was trained 10 times, each on a unique split of the data. The final model achieved an average area under the curve for the receiver operating characteristic curve of 0.7717 (95% CI±0.0052), calibration curve analysis revealed a mean slope of 1.243 (95% CI±0.075) and intercept of −0.073 (95% CI±0.034), indicating a good correlation between prediction and ground truth. Post-classification, tuned for F1 score, accuracy was 71.43% (95% CI±1.28), with an F1 score of 0.5642 (95% CI±0.0052) and a negative predictive value of 0.8660 (95% CI±0.0048), respectively. The ResNet displayed comparable or surpassing metrics to prior presented models.
Conclusion
The model exhibited clinically meaningful performance, notably its high negative predictive accuracy. Therefore, clinical use of comparable neural networks in first-line, quick-response triage of patients with chest pain or dyspnoea appears as a valuable option in future medical practice.
Keywords: Coronary Artery Disease, Acute Coronary Syndrome, Chest Pain
WHAT IS ALREADY KNOWN ON THIS TOPIC
Rapid and accurate diagnosis of myocardial infarction (MI) is essential for patient outcomes. ECGs have been analysed to receive information about relevant biomarkers, such as troponin, using deep learning models, however, with low negative predictive accuracy of under 80% and not taking into consideration sex and assay-specific cut-offs.
WHAT THIS STUDY ADDS
This study presents a deep learning model, trained on a small but diverse dataset from multiple centres, to predict troponin elevation in emergency room patients with chest pain based on an ECG. The model achieved high negative predictive accuracy of 86%, with data that take into account the assay-specific cut-offs used in different laboratories and for sex differences.
HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY
Models such as the one presented may provide an opportunity, if implemented clinically, to provide an accurate troponin information in a triage setting before laboratory analysis is possible, even before full cardiological expertise reaches the patient. Adding such information earlier on may aid medical personnel in decision-making concerning MI diagnoses.
Introduction
The ECG is fundamental for the assessment of patients with chest pain. The use of artificial intelligence (AI) might widen the clinical use of ECG to detect abnormalities hardly visual to the human eye, in addition to improving diagnostic precision of visible ECG changes. Neural networks have been applied for processing of ECG signals for the last decades,1 and over the past years multiple deep learning (DL) algorithms have been presented, some with astounding accuracy.2 Currently, there exist algorithms for arrhythmia classification, myocardial infarction (MI) detection and prediction of biomarkers.3,5 A common approach are convolutional neural networks (CNN), which have demonstrated an extraordinary ability to interpret ECGs.6
When diagnosing ST-elevation MI (STEMI), using an ECG, trained physicians have a sensitivity of 65% and area under the curve (AUC) for the receiver operating characteristic (ROC) of 0.72.7 This diagnosis is of course aided in the clinical setting by a combination of other parameters, particularly symptoms and haemodynamic status. However, there remains room for improvement in the clinical detection of MI, especially with rapid tools requiring less human input. Especially as delayed diagnoses, and thereby delayed treatment, can have severe adverse effects on patient outcomes.8
Currently, elevation of troponin (Tn) is a criterion for the diagnosis of MI.9 The analysis, however, takes around 45–60 min from blood sampling to result.10 Furthermore, Tn biomarkers are not immediately elevated on cardiac necrosis, requiring testing at multiple time points to determine whether elevation occurs and to determine whether or not Tn levels are dynamic.11 These factors might lead to a delay in patient treatment.
With the rapid development of DL algorithms, an interesting opportunity opens up to extract information from ECGs and instantly provide an AI interpretation to physicians and medical staff during triage, to more quickly identify patients with MI. This has been attempted recently with a large dataset of 64 728 ECGs, achieving an AUC of 0.783 (95% CI 0.780 to 0.786).12 This model does, however, only use one cut-off of 20 ng/L for patients’ Tn values and does not take into consideration the modern existing sex differences in Tn diagnostics. Therefore, we aimed at training a neural network for Tn prediction from an initial 10 s, 12-lead ECG taken at triage in patients with chest pain or dyspnoea.
Methods
Patient population
Data were collected retrospectively. The population included all patients who received a 12-lead ECG examination within 1 hour of admission to the emergency room (ER) at two hospitals in the Swedish region of Västra Götaland, for chest pain or dyspnoea, between 1 January 2015 and 30 June 2023, and had a high-sensitivity Tn test done within 6 hours of ER admission. The dataset does not distinguish between non-STEMI and STEMI.
Preprocessing of ECGs
For each patient, the first ECG of the ER visits was used for further analysis, that is, the ECG taken in triage on arrival to the hospital. Triage was performed by a nurse on patient arrival at the ER, during which it is standard procedure for patients with chest pain or dyspnoea to undergo a standard 12-lead ECG. No ECGs obtained out of hospital (eg, ambulance) were included in this study, as the purpose of the to-be-trained neural networks was to be used on ECGs recorded during triage. ECGs by nature can vary greatly in amplitude, between persons and between recording devices; for this reason normalisation is a common method of accounting for variations while not disturbing the critical time series aspect of the data.13 Furthermore, CNN models use a process known as batch normalisation between convolutional layers as these layers struggle with input of varying magnitude.14 For this reason, normalisation of the input into the first layer of the CNN can be performed to achieve better performance and quicker convergence during training. Normalisation is often performed with the standard score method which pushes the mean of the ECG to 0 and the SD to 1.15
Neural networks were trained employing ECG input as one-dimensional data, meaning the ECG is not represented as an image but rather as values which, if plotted, would make up the classic ECG image. This removes the majority of the blank image space of an ECG and also removes other non-data factors such as line thickness and grid colour, with the additional advantage of taking less computing power as the size of the input is drastically reduced.
Tn analyses
The usage of different Tn analytical methods at the clinical sites/over the duration of data acquisition necessitates usage of multiple unique cut-offs when determining elevation; all cut-offs represent the 99th percentile for the used assay. For TnT, a cut-off of 14 ng/L was used. For TnI, sex and laboratory-specific cut-offs were used, 16 ng/L for women and 35 ng/L for men analysed in one laboratory; the other laboratory applied the cut-offs of 36 ng/L for women and 56 ng/L for men. (For more in-depth information on analyses, see online supplemental material 1.) We used these cut-off values as they accurately represent those used in the clinical process for these assays. This discrepancy stems from the definition of Tn elevation in the definition of MI, which states elevation is above the 99th percentile in the normal population,9 leading to each laboratory calibrating their assay according to the instructions of the manufacturer. This difference in assay and calibration leads to the absolute variation in cut-offs, since these all represent the 99th percentile values become comparable across assays. The comparability has been further explored in online supplemental material 2. Each Tn value was compared with its cut-off to determine a binary variable of elevation which was either true or false. This allows for the most accurate representation of a clinically relevant Tn elevation, as it uses the level which is considered relevant in the clinic. If a patient had multiple Tn tests done, if any one was elevated, the patient was considered having Tn elevation.
DL model development and training
A multitude of model approaches were tested during exploratory model evaluation, for example, Xtreme Gradient Boosting, two different residual network architectures, as well as a fusion model of two residual networks. The final model was selected based on performance metrics during training iterations, aiming for a maximum possible AUC of the ROC curve, as well as maximising the F1 score. Learning rates were tested between 0.1 and 0.000001, decreasing every step to 0.1 times the previous learning rate.
The input data were split into three parts in a 70-15-15 split, training, validation and test sets. The model was written in PyTorch and trained on an NVIDIA RTX-4090 GPU. There was no constant number of epochs, instead a minimum number of epochs were set to 10 and early stopping was implemented if the validation loss did not improve over the last 10 epochs.
Model evaluation and statistical analysis
For the purpose of evaluation, each model was trained 10 times, each time on a different split of the same inputs. This ensures that the data collected are not dependent on the input split, but rather are consistent over different input data. The models were initialised with a constant seed, which ensures constant starting weights over trials. Evaluation was also performed using a calibration curve, which describes the models’ predictive performance within ranges of predicted probability called bins. It allows for a threshold independent evaluation and a visualisation of how close to accurate the predictive performance is within each bin. A perfect curve would be described by a slope of 1 and intercept of 0, as each predicted probability is exactly correlated with the actual probability. Furthermore, machine dichotomisation was performed; a threshold is applied to the model output, a likelihood prediction between 0 and 1, changing the output into a binary outcome, elevated or non-elevated Tn. This threshold was selected for each model at the point where it achieves the best F1 score, as this does not eliminate any possible metrics from further evaluation, whereas if optimised for a set metric such as sensitivity, sensitivity can no longer be used to evaluate the models. This is due to the fact that the F1 score is the harmonic mean between positive predictive value (PPV) and sensitivity, and a model optimised for it strikes a balance between both of these scores. At this threshold, evaluation was performed using sensitivity, accuracy, positive likelihood ratio (PLR), negative likelihood ratio (NLR), negative predictive value (NPV) and PPV. For all of these metrics a 95% CI was calculated.
Qualitative evaluation
Saliency maps were computed on one true sample and one false sample from the test set. These samples were chosen prior to saliency map computation, based on their similar ECG morphology despite belonging to different classes, the lack of arrhythmia and absence of noise and artefacts. The saliency maps were calculated by computing the gradient of every input sample, normalising each leads gradients to between 0 and 1, and then the gradients were overlaid onto the ECG sample to visualise the segments used by the model to make the prediction.
During evaluation, a focus was placed on the common ischaemic changes in ECGs, with a focus on the ST-segment. Furthermore, a point of interest is expected just before the QRS complex to determine the expected location of the J-point after the QRS complex. However, this task is not currently being performed by physicians and therefore there are no known criteria to determine Tn elevation from ECGs as there are for other ECG use cases such as arrhythmia diagnosis; meaning, that our expectation remains speculation, and there may be important information in other portions of the ECG.
Results
During the study period, a total of 111 094 visits were made to the ERs for the selected symptoms; of these visits, 15 856 were matched with a high-sensitivity Tn value and an ECG, the exclusion procedure can be seen in figure 1. Of the final subjects, 4148 (26.2%) had at least one elevated Tn during their hospital stay; of those, 2433 were male and 1715 were female. The study included 11 708 (73.8%) patients without elevated Tn; of those, 6402 were male and 5306 were female.
Figure 1. Dataset population. Displaying the filtering of emergency room (ER) visits with chest pain or dyspnoea down to the final population.
ECGs were normalised by using the standard score method and normalising to a mean of 0 and an SD of 0.5. The original units were in millivolts (mV), and the R-peaks could vary between 60 and 2000 mV between leads and patients.
The final residual convolutional neural network (ResNet) model (for design details, cf online supplemental material 3), based on threshold independent evaluation of the ROC, achieved an average AUC of 0.7717 (95% CI±0.0052), as given in figure 2. The ROC curve exhibits a tight 95% CI with the maximum deviation at any given false positive rate value being ±0.0274. For further threshold independent evaluation the calibration curve with 10 bins (ie, the model outcome likelihood predictions grouped into 10 equally large groups) has been calculated, as given in figure 3, displaying the mean of the values in each bin over the runs. The curves of all models yield a mean slope of 1.243 (95% CI±0.075), with an intercept of −0.073 (95% CI±0.034).
Figure 2. Receiver operating characteristic (ROC) curve of the over 10 models of the same architecture (mean±95% CI), each trained on a unique split of the data into testing and training data, for the final neural network predicting likelihood of high-sensitivity troponin elevation from 12-lead ECG obtained from patients presenting to an emergency room with chest pain or dyspnoea. The area under the curve (AUC) of 0.7717±0.0052 (mean±95% CI) indicates adequate prognostic power, as the AUC under the line of identity (0.5) represents full randomness of prediction, that is, zero predictive power of the model.
Figure 3. Calibration curve over 10 models of the same architecture (mean±95% CI), each trained on a unique split of the data into testing and training data, for the final neural network predicting likelihood of high-sensitivity troponin elevation from 12-lead ECG obtained from patients presenting to an emergency room with chest pain or dyspnoea.
When evaluating the models post-dichotomisation, tuned for a maximum F1 score and then averaged, adequate model performance was confirmed using accuracy, F1 score, sensitivity, PPV and NPV, as displayed in figure 4. Moreover, a PLR of 2.5269±0.0032 and an NLR of 0.4251±0.0032, respectively, were achieved. The NPV of 0.866 in the model indicates that above 86.6% of all negative predictions were correct. Importantly, if the model were to predict with completely random accuracy, then this figure would be 73.8% as that is the fraction of negatives in the population. PPV was even more compelling, reaching 47.8% in contrast to the 26.2% fraction of positives in the population.
Figure 4. Mean evaluation metrics over 10 models of the same architecture (mean±95% CI), each trained on a unique split of the data into testing and training data. All metrics should ideally approach 1. Due to imbalanced class sizes, negative predictive value (NPV) baseline, if completely random, is 0.738, and positive predictive value (PPV) is 0.262. AUC, area under the curve.
The final ResNet model, when presented with a ‘true sample’ (ie, a sample with an elevated Tn), yields a result extracting a segment prior to the QRS complex and ST-segment, specifically the J-point, as can be seen in figure 5A (saliency map). The ‘false sample’ (ie, a sample with non-elevated Tn) yields a wider feature extract, all the way from the T-wave until the next QRS complex, however, with a focus being on the TP-segment, indicating that the classification as negative originates from a lack of clearly defined features leading to larger activation patterns. There are also features extracted from the ST-segment in the false sample, as can be seen in figure 5B.
Figure 5. Saliency maps of the residual convolutional neural network (ResNet) model showing its regions of interest, taken from the test set, the leads displayed are V1–6. Examples of the ‘true class’ representing an elevated troponin (A) and the ‘false class’ representing a non-elevated troponin (B).
Discussion
We created a CNN model capable of predicting Tn elevation from ER triage 12-lead ECGs, achieving metrics comparable to or surpassing those of the best prior published model for the task, with one-quarter of the dataset of the prior model, that is, 15 856 ECGs for the model presented here versus 64 728 ECGs.12 In all relevant characteristics, such as AUC under the ROC, NPV, PLR and NLR, the new model achieved comparable or better metrics, that is, 0.77, 0.86, 2.53 and 0.43, respectively, compared with 0.78, 0.77, 2.1 and 0.32 of the model published in 202312 . Moreover, the model extracts features from the patients’ ECG (depicted in saliency maps), which are in line with those deemed relevant by clinicians when determining MI status of patients, such as the J-point and ST-segment. The model’s NPV at 86.6% is considerably greater than the fraction of negatives, which allows for the conclusion that the model’s performance can, with great accuracy, filter out negative samples. While the PPV only reaches 47.6%, this is significantly greater than the fraction of positives in the dataset. The combination of these metrics allows for the conclusion that the model can truly detect Tn elevation in many cases, while not detecting all elevations it shows that with further optimisation and larger datasets this task may be solvable.
The model presented here is the first for this task that has been trained on sex-specific TnI thresholds allowing for an adequate sex-specific use in diagnostics, even without feeding the model patient sex as a variable. This will allow for more accurate diagnoses of Tn elevation in especially women as they are not fully represented in the older models which implement a hard cut-off at 20 ng/L, not taking into account for intersex or intercentre variability.
The shape, slope and intercept of the calibration curve indicate mild overdiagnosis for model predictions of high likelihood of Tn elevation, as well as mild underprediction for predictions of a low likelihood of Tn elevation, switching between these two characteristics at a probability of around 40%. This performance opens up for the possibility of avoiding machine dichotomisation and simply displaying the predicted probability of Tn elevation to the ECG interpreter, returning more detailed information back to the clinical personnel who can incorporate it into their broader overview of the patient.
As compared with published data, our model appears to outperform clinicians working in the ER, with regard to the task of STEMI diagnosis, where clinicians reach a sensitivity of 0.65 and AUC of 0.72.7 This comparison is indirect (ie, our model does not aim to diagnose STEMI from ECGs but rather the Tn elevation), however, the model presented here could, in the STEMI cases missed by the physicians, aid them in reconsidering their diagnosis. Furthermore, implementation of this system in triage can aid nursing staff in deciding when to alert physicians to possible MI, to ensure that patients receive care in a timely manner.
This work has some limitations, among which is the small amount of 15 856 ECGs available at the time of model building. More ECGs and Tn data are currently being gathered from more clinical centres yielding a broader population and larger sample size which may further enhance model performance. A further expansion of the training dataset may also increase the ability to generalise on broader populations as more variance is introduced.
Furthermore, there is also a lack of preselection of ECGs in the current dataset used for model training. While this does yield a broad input aiming to make the model applicable to as broad a population as possible, it comes with certain drawbacks. The main issue that arises is the lack of differentiation between Tn elevations that are caused by type I MI and those caused by type II MI. The issue presented by type II MI is their lack of clear ECG changes, with only 3–24% having ST-elevation.9 Furthermore, other ischaemic ECG changes are also significantly less common than in type I MI.16 Moreover, we did not remove patients with arrhythmia, previous MI or branch bundle blocks. Although these diagnoses may render the interpretation of an ECG more complex, they are abundant in a population presenting with chest pain, and the DL model is capable of detecting patterns also among these subgroups, provided that the training data are large enough.
Additionally, the data were not standardised to require a uniform regimen of Tn testing, beyond the requirement that first test was to be done within 6 hours, as this would be unfeasible with retrospective data collected from ERs. Ideally all patients would receive multiple Tn tests at set time intervals to achieve a uniform dataset to determine Tn elevation with absolute certainty. In the current clinical practice of sending home patients without repeat Tn testing if a clinician rules out cardiac injury, it is possible that elevation was missed due to clinician error, however, we believe that the size of the dataset prevents these outlier cases from affecting model performance.
A common method of overcoming the limitations in ECG dataset size for model building is to use ‘data augmentation’, that is, either creating synthetic ECGs based on existing data or applying transformations onto ECGs and introducing these as ‘new’ into the dataset.17,19 A systematic review of data augmentation of ECG signals for AI applications concluded that no data augmentation method is applicable across all types and tasks of DL models, as many methods relying on the transformation of existing ECGs can, for example, eliminate important temporal information and resolution during transformation.20 Therefore, we decided against data augmentation to ensure that our model was trained only on real-world data, especially as it is unclear which features of the ECG are relevant for detection of Tn elevation.
Conclusion
This study has demonstrated that a neural network AI algorithm is capable of predicting high-sensitivity Tn elevation with NPV of 86.6% from 12-lead ECGs taken in ER patients presenting with chest pain or dyspnoea. The usage of individual cut-offs for sex ensures adequate diagnostic power in both male and female patient populations. Continued model evolution using larger datasets, allowing for more homogeneous subgroups for algorithm training in the future, may yield further improved diagnostic capabilities in the future.
supplementary material
Acknowledgements
In this project, ChatGPT has been used to assist in and accelerate coding and debugging python code for model creation, all of the generated code and suggestions have been checked for functionality and correctness by the corresponding author. No other generative AI has been used during the process of the research or writing of this work.
Footnotes
Funding: This study was funded by the Swedish Society for Medicine (Svenska Läkarsällskapet, grant number SLS1000926), the Swedish state under the agreement between the Swedish government and the county councils, the ALF agreement (ALFGBG-998297) and the Wallenberg Centre for Molecular and Translational Medicine (WCMTM).
Provenance and peer review: Not commissioned; externally peer reviewed.
Patient consent for publication: Not applicable.
Ethics approval: This study involves human participants and was approved by the Swedish Ethical Review Authority (Dnr 2023-03921-01). The ethical review board agreed that patient consent would not be necessary as the patient data were stored securely and not evaluated on an individual basis but rather anonymised and used to train the models.
Data availability free text: The dataset cannot be made available publicly as it is identifiable as the Swedish personal number was used to match data from the laboratory, emergency room and ECG database.
Contributor Information
Lukas Hilgendorf, Email: lukas.hilgendorf@gu.se.
Petur Petursson, Email: petur.petursson@vgregion.se.
Vibha Gupta, Email: vibha.gupta@gu.se.
Truls Ramunddal, Email: truls.ramunddal@vgregion.se.
Erik Andersson, Email: gusanderdm@student.gu.se.
Peter Lundgren, Email: peter.lundgren@vgregion.se.
Christian Dworeck, Email: christian.dworeck@vgregion.se.
Charlotta Ljungman, Email: charlotta.ljungman@vgregion.se.
Jan Boren, Email: jan.boren@wlab.gu.se.
Aidin Rawshani, Email: aidin.rawshani@gu.se.
Elmir Omerovic, Email: elmir@wlab.gu.se.
Gustav Smith, Email: gustav.smith@med.lu.se.
Zacharias Mandalenakis, Email: zacharias.mandalenakis@vgregion.se.
Kristofer Skoglund, Email: kristofer.skoglund@gu.se.
Araz Rawshani, Email: araz.rawshani@vgregion.se.
Data availability statement
No data are available.
references
- 1.Miller AS, Blott BH, Hames TK. Review of neural network applications in medical imaging and signal processing. Med Biol Eng Comput. 1992;30:449–64. doi: 10.1007/BF02457822. [DOI] [PubMed] [Google Scholar]
- 2.Sakli N, Ghabri H, Soufiene BO, et al. ResNet-50 for 12-Lead Electrocardiogram Automated Diagnosis. Comput Intell Neurosci. 2022;2022:7617551. doi: 10.1155/2022/7617551. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Singh S, Chaudhary R, Bliden KP, et al. Meta-Analysis of the Performance of AI-Driven ECG Interpretation in the Diagnosis of Valvular Heart Diseases. Am J Cardiol. 2024;213:126–31. doi: 10.1016/j.amjcard.2023.12.015. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Siontis KC, Noseworthy PA, Attia ZI, et al. Artificial intelligence-enhanced electrocardiography in cardiovascular disease management. Nat Rev Cardiol. 2021;18:465–78. doi: 10.1038/s41569-020-00503-2. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Cho Y, Kwon J, Kim K-H, et al. Artificial intelligence algorithm for detecting myocardial infarction using six-lead electrocardiography. Sci Rep. 2020;10:20495. doi: 10.1038/s41598-020-77599-6. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Data Basecamp What are convolutional neural networks (CNN)? [22-Jan-2021]. https://databasecamp.de/en/ml/convolutional-neural-networks Available. Accessed.
- 7.McCabe JM, Armstrong EJ, Ku I, et al. Physician accuracy in interpreting potential ST-segment elevation myocardial infarction electrocardiograms. J Am Heart Assoc. 2013;2:e000268. doi: 10.1161/JAHA.113.000268. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Braunwald E. Evolution of the management of acute myocardial infarction: a 20th century saga. The Lancet. 1998;352:1771–4. doi: 10.1016/S0140-6736(98)03212-7. [DOI] [PubMed] [Google Scholar]
- 9.Thygesen K, Alpert JS, Jaffe AS, et al. Fourth universal definition of myocardial infarction (2018) Eur Heart J. 2019;40:237–69. doi: 10.1093/eurheartj/ehy462. [DOI] [PubMed] [Google Scholar]
- 10.Cardiac troponin. [8-Apr-2024]. https://www.siemens-healthineers.com/laboratory-diagnostics/assays-by-diseases-conditions/cardiac-assays/cardiac-troponin-assays Available. Accessed.
- 11.Morrow DA, Cannon CP, Jesse RL, et al. National Academy of Clinical Biochemistry Laboratory Medicine Practice Guidelines: clinical characteristics and utilization of biochemical markers in acute coronary syndromes. Clin Chem. 2007;53:552–74. doi: 10.1373/clinchem.2006.084194. [DOI] [PubMed] [Google Scholar]
- 12.Chaudhari GR, Mayfield JJ, Barrios JP, et al. Deep learning augmented ECG analysis to identify biomarker-defined myocardial injury. Sci Rep. 2023;13:3364. doi: 10.1038/s41598-023-29989-9. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Salimi A, Kalmady SV, Hindle A, et al. arXiv; 2023. [4-Mar-2024]. Exploring best practices for ecg signal processing in machine learning.http://arxiv.org/abs/2311.04229 Available. Accessed. [Google Scholar]
- 14.Ioffe S, Normalization SCB. Accelerating Deep Network Training by Reducing Internal Covariate Shift. arXiv; 2015. [11-Apr-2024]. http://arxiv.org/abs/1502.03167 Available. Accessed. [Google Scholar]
- 15.SOP-z-score-standard-score-calculation.pdf. [12-Apr-2024]. https://hiv-forschung.de/wp-content/uploads/SOP-Z-score-standard-score-calculation.pdf Available. Accessed.
- 16.Merlo AC, Bona RD, Ameri P, et al. Type 2 myocardial infarction: a diagnostic and therapeutic challenge in contemporary cardiology. Intern Emerg Med. 2022;17:317–24. doi: 10.1007/s11739-021-02920-8. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17.Zhang P, Cheng J, Zhao Y. In: Artificial Intelligence and Security. Sun X, Wang J, Bertino E, editors. Singapore: Springer; 2020. Classification of ecg signals based on lstm and cnn; pp. 278–89. [Google Scholar]
- 18.Mehari T, Strodthoff N. Self-supervised representation learning from 12-lead ECG data. Comput Biol Med. 2022 doi: 10.1016/j.compbiomed.2021.105114. [DOI] [PubMed] [Google Scholar]
- 19.Liu Y, Xie H, Cao Q, et al. Multi-label classification of multi-lead ecg based on deep 1d convolutional neural networks with residual and attention mechanism: 2021 computing in cardiology (cinc) [14-Aug-2024]. https://ieeexplore.ieee.org/abstract/document/9662873 Available. Accessed.
- 20.Rahman MM, Rivolta MW, Badilini F, et al. A Systematic Survey of Data Augmentation of ECG Signals for AI Applications. Sensors (Basel) 2023;23:5237. doi: 10.3390/s23115237. [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
No data are available.