Comprehensive survey of computational ECG analysis: Databases,methods and applications

1、Learning algorithms

classifiers（most common and highest-performing):

1. Support Vector Machines（SVM 支持向量机）(Qin et al., 2017; Raj & Ray, 2018)  ( Qin, Q., Li, J., Zhang, L., Yue, Y., & Liu, C. (2017). Combining Low dimensional Wavelet Features and Support Vector Machine for Arrhyth mia Beat Classifification. Scientifific Reports, 7 , Article 6067. doi:10.1038/s41598-017-06596-z.)    ( Raj, S., & Ray, K. (2018). A Personalized Arrhythmia Monitoring Platform.1480 Scientifific Reports, 8 , Article 11395. doi:10.1038/s41598-018-29690-2.)
2. Random Forest (RF 随机森林) (Lyon et al., 2018)  (Lyon, A., Minchol´e, A., Mart´ınez, J. P., Laguna, P., & Rodriguez, B. (2018).Computational techniques for ECG analysis and interpretation in light of theircontribution to medical advances. Journal of The Royal Society Interface, 15 ,Article 20170821. doi:10.1098/rsif.2017.0821.)
3. Nearest Neighbors Classififier (邻近算法) (Sprager et al.,2017)
4. MLP networks (Multi-layer Perception 多层感知机) (Chromik et al., 2021; Li et al., 2017). ( Chromik, J., Pirl, L., Beilharz, J., Arnrich, B., & Polze, A. (2021). Certainty in

   QRS detection with artifificial neural networks. Biomedical Signal Processing and Control, 68 , Article 102628. doi:10.1016/j.bspc.2021.102628.)

stochastic optimization methods( an important step towards achieving best possible performance in the end-task):

1. ABC (Adaboost Classifiers) (Raj & Ray, 2018) (Raj, S., & Ray, K. (2018). A Personalized Arrhythmia Monitoring Platform. Scientifific Reports, 8 , Article 11395. doi:10.1038/s41598-018-29690-2.)
2. GA (Li et al., 2017)（Li, H., Yuan, D., Ma, X., Cui, D., & Cao, L. (2017). Genetic algorithm for the optimization of features and neural networks in ECG signals classifification. Scientifific Reports, 7 , Article 41011. doi:10.1038/srep41011）

Deep Neural Networks:

1. Convolutional neural networks(CNN) (Kiranyaz et al., 2017; Smith et al., 2019b) （

   Kiranyaz, S., Ince, T., & Gabbouj, M. (2017). Personalized Monitoring and Advance Warning System for Cardiac Arrhythmias. Scientifific Reports, 7 , Article 9270. doi:10.1038/s41598-017-09544-z.） （ Smith, S. W., Walsh, B., Grauer, K., Wang, K., Rapin, J., Li, J., Fennell, W., & Taboulet, P. (2019b). A deep neural network learning algorithm outperforms a conventional algorithm for emergency department electrocardiogram interpretation. Journal of Electrocardiology, 52 , 88–95.1530 doi:https://doi.org/10.1016/j.jelectrocard.2018.11.013.）

2. Residual neural networks(ResNet) (Hannun et al., 2019; Ribeiro et al., 2020) （Hannun, A., Rajpurkar, P., Haghpanahi, M., Tison, G., Bourn, C., Turakhia,M., & Ng, A. (2019). Cardiologist-level arrhythmia detection and classifification in ambulatory electrocardiograms using a deep neural network. Nature Medicine, 25 , 65–69. doi:10.1038/s41591-018-0268-3.）（ Ribeiro, A. H., Ribeiro, M., Paix˜ao, G., Oliveira, D., Gomes, P., Canazart, J., Ferreira, M., Andersson, C., Macfarlane, P., Wagner, M., Sch¨on, T., & Ribeiro, A. L. (2020). Automatic diagnosis of the 12-lead ECG using a deep neural network. Nature Communications, 11 , Article 1760. doi:10.1038/s41467-020-15432-4.）
3. Long Short-Term Memory (LSTM) (Chauhan & Vig, 2015; Yao et al., 2020) （Chauhan, S., & Vig, L. (2015). Anomaly detection in ECG time signals via deep long short-term memory networks. In 2015 IEEE International Conference on Data Science and Advanced Analytics (DSAA) (pp. 1–7). doi:10.1109/DSAA.2015.7344872. ）（Yao, Q., Wang, R., Fan, X., Liu, J., & Li, Y. (2020). Multi-class Arrhythmia detection from 12-lead varied-length ECG using Attention-based Time Incremental Convolutional Neural Network. Information Fusion, 53 , 174–182. doi:10.1016/j.inffus.2019.06.024.）
4. Gated Recurrent Units (GRU) (Chen et al., 2020) （Chen, T.-M., Huang, C.-H., Shih, E. S., Hu, Y.-F., & Hwang, M.-J. (2020). Detection and Classifification of Cardiac Arrhythmias by a Challenge-Best Deep Learning Neural Network Model. iScience, 23 , Article 100886. doi:10.1016/j.isci.2020.100886.）
5. Attention mechanisms (Chen et al., 2020; Hong et al., 2019; Mousavi et al., 2019; Yaoet al., 2020).（Chen, T.-M., Huang, C.-H., Shih, E. S., Hu, Y.-F., & Hwang, M.-J. (2020). Detection and Classifification of Cardiac Arrhythmias by a Challenge-Best Deep Learning Neural Network Model. iScience, 23 , Article 100886. doi:10.1016/j.isci.2020.100886.）（Hong, S., Xiao, C., Ma, T., Li, H., & Sun, J. (2019). MINA: Multilevel Knowledge-Guided Attention for Modeling Electrocardiography Signals. In Proceedings of the Twenty-Eighth International Joint Conference on Artifificial Intelligence, IJCAI-19 (pp. 5888–5894). doi:10.24963/ijcai.2019/816.）（Mousavi, S., Afghah, F., Razi, A., & Acharya, U. R. (2019). ECGNET: Learning Where to Attend for Detection of Atrial Fibrillation with Deep Visual Attention. In 2019 IEEE EMBS International Conference on Biomedical Health Informatics (BHI) (pp. 1–4). doi:10.1109/BHI.2019.8834637.）（Yao, Q., Wang, R., Fan, X., Liu, J., & Li, Y. (2020). Multi-class Arrhythmia detection from 12-lead varied-length ECG using Attention-based TimeIncremental Convolutional Neural Network. Information Fusion, 53 , 174–182.doi:10.1016/j.inffus.2019.06.024.）
6. attention mechanism and LSTM units(Mousavi et al.,2019) （Mousavi, S., Afghah, F., Razi, A., & Acharya, U. R. (2019). ECGNET: Learning Where to Attend for Detection of Atrial Fibrillation with Deep Visual Attention. In 2019 IEEE EMBS International Conference on Biomedical Health Informatics (BHI) (pp. 1–4). doi:10.1109/BHI.2019.8834637.）
7. Encoder-decoder architectures for arrhthmia detection (Mousavi & Afghah, 2019)  and fetal ECG signal denoising(Fotiadou et al., 2020)（Mousavi, S., & Afghah, F. (2019). Inter- and Intra- Patient ECG Heart beat Classifification for Arrhythmia Detection: A Sequence to Sequence Deep Learning Approach. In ICASSP 2019 - 2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (pp. 1308–1312).1445 doi:10.1109/ICASSP.2019.8683140）（Fotiadou, E., Konopczy´nski, T., Hesser, J., & Vullings, R. (2020). End-toend trained encoder-decoder convolutional neural network for fetal electrocardiogram signal denoising. Physiological Measurement, 41 , Article 015005.1300 doi:10.1088/1361-6579/ab69b9.）
8. feature extractors combined with simple classififiers such as 1-Nearest Neighbour
   (Labati et al., 2018)（Labati, R., Mu˜noz, E., Piuri, V., Sassi, R., & Scotti, F. (2018). Deep-ECG: Convolutional Neural Networks for ECG biometric recognition. Pattern Recognition Letters, 126 , 78–85. doi:10.1016/j.patrec.2018.03.028.）

 ECG Databases：（部分）

including sampling frequency, number of leads, length and size (in terms of number of recordings); annotations included (in terms of targeted application areas); and notable studies using them.

 

Most of the databases listed in the table are available through the PhysioNet repository (https://physionet.org/)  , which is an online platform for sharing medical data.

Additionally, a few of the datasets can be found on other public repositories such as fifigshare (https://fifigshare.com/) , Zenodo ( https://zenodo.org/ ) and IEEE Data Port ( https://ieee-dataport.org/).

2、Applications

1. morphological and rhythmic arrhythmia detection
2. signal quality assessment
3. biometric identifification
4. respiration estimation
5. fetal ECG extraction
6. physical and emotional monitoring

2.1 Heartbeat classification for arrythmia detection:

(including representative heartbeats from a variety of data sourse, instead of only one database)

two mian categories:

1. ***morphological(characterized by the irregularity of a single heartbeat. )
2.  rhythmic arrthmias(characterized by a set of irregular heartbeats.)

database:

The most popular publicly available arrhythmia database is the MIT-BIH Arrhythmia database (Moody & Mark, 2001).(perfect results: overall precision and recall of around 96-97%.( Acharya et al., 2017)（Moody, G., & Mark, R. (1983). A new method for detecting atrial fifibrillation using R-R intervals. Computers in Cardiology, 10 , 227–230. doi:10.1093/

1435 europace/eum096. ）（Acharya, U. R., Oh, S. L., Hagiwara, Y., Tan, J. H., Adam, M., Gertych, A.,& Tan, R. S. (2017). A deep convolutional neural network model to classify heartbeats. Computers in Biology and Medicine, 89 , 389–396. doi:10.1016/j.compbiomed.2017.08.022.）

a few novel datasets that contain heartbeat form labels( not been used as extensively as the MIT-BIH Arrhythmia Detection database):

1. PTB-XL database (Wagner et al.,2020) （Wagner, P., Strodtho↵, N., Bousseljot, R., Kreiseler, D., Lunze, F., Samek,W., & Schae↵ter, T. (2020). PTB-XL: A Large Publicly Available electrocardiography Dataset. Scientifific Data, 7 , Article 154. doi:10.1038/

   s41597-020-0495-6）

2. the Shaoxing People’s Hospital’s 10,000 patients arrhythmia database(Zheng et al., 2020)（Zheng, J., Zhang, J., Danioko, S., Yao, H., Guo, H., & Rakovski, C. (2020). A 12-lead Electrocardiogram Database for Arrhythmia Research Covering More Than 10,000 Patients. Scientifific Data, 7 , Article 48. doi:10.1038/s41597-020-0386-x）

(this new generation of arrhythmia datasets provides a set of labels for each ECG recording.)

(The recordings in these datasets are much shorter, usually around 10 seconds.)

many new interesting datasets containing 12-lead ECG recordings for utilizing all 12 leads present in standard clinic ECG:

1. 2018 China Physiological Signal Challenge Dataset (Liu et al., 2018)（Liu, F., Liu, C., Zhao, L., Zhang, X., Wu, X., Xu, X., Liu, Y., Ma, C., Wei,S., He, Z., Li, J., & Ng, E. (2018). An Open Access Database for Evaluatingthe Algorithms of Electrocardiogram Rhythm and Morphology AbnormalityDetection. Journal of Medical Imaging and Health Informatics, 8 , 1368–1373.1395 doi:10.1166/jmihi.2018.2442）
2. the smaller but highly annotated Lobachevsky University Electrocardiography Database (LUDB) (Kalyakulina et al., 2020) with 200 recordings（Kalyakulina, A. I., Yusipov, I. I., Moskalenko, V. A., Nikolskiy, A. V., Kosonogov, K. A., Osipov, G. V., Zolotykh, N. Y., & Ivanchenko, M. V. (2020).LUDB: A New Open-Access Validation Tool for Electrocardiogram Delineation Algorithms. IEEE Access, 8 , 186181–186190. doi:10.1109/ACCESS.2020.3029211.）
3. the previously mentioned PTB-XL database
4. the Shaoxing People’s Hospital’s 10,000 patients arrhythmia database

These recordings are generally ranging from 10 seconds up to 1 minute in most databases.

all 12 ECG leads 

deep learrning:

the latest best performing methods:

1. the study in Mousavi & Afghah (2019), which uses a combination of a simple convolutional architecture with 3 layers, followed by an encoder-decoder architecture, built from bidirectional LSTM units.（Mousavi, S., & Afghah, F. (2019). Inter- and Intra- Patient ECG Heartbeat Classifification for Arrhythmia Detection: A Sequence to Sequence Deep Learning Approach. In ICASSP 2019 - 2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (pp. 1308–1312). doi:10.1109/ICASSP.2019.8683140）
2. SMOTE technique (Sellami & Hwang,2018), where a novel loss function for a convolutional neural network is proposed.(changes the loss weights dynamically according to the distribution of classes in each batch, is referred to as batch-weighted loss.)(F1-score:0.897)（Sellami, A., & Hwang, H. (2018). A Robust Deep Convolutional Neural Network with Batch-Weighted Loss for Heartbeat Classifification. Expert Systems withApplications, 122 , 75–84. doi:10.1016/j.eswa.2018.12.037.）
3. improved residual neural network (xresnet)(Strodtho et al. (2021)(AUC:0.896)（Strodtho, N., Wagner, P., Schaeter, T., & Samek, W. (2021). Deep Learning for ECG Analysis: Benchmarks and Insights from PTB-XL. IEEE Journal of Biomedical and Health Informatics, 25 , 1519–1528. doi:10.1109/JBHI.2020.3022989）
4. attention-based time-incremental convolutional neural network that successfully fuses the information from all leads, both spatially, using convolutional layers, and temporally, using LSTM units. (Yao et al., 2020)（Yao, Q., Wang, R., Fan, X., Liu, J., & Li, Y. (2020). Multi-class Arrhythmia detection from 12-lead varied-length ECG using Attention-based TimeIncremental Convolutional Neural Network. Information Fusion, 53 , 174–182. doi:10.1016/j.inffus.2019.06.024）
5. an interesting classifification approach to utilize all 12 leads, which is also based on convolutional neural networks, but is enhanced with an ensemble model combining 12-and 1-lead models(Chen et al., 2020)（Chen, T.-M., Huang, C.-H., Shih, E. S., Hu, Y.-F., & Hwang, M.-J. (2020). Detection and Classifification of Cardiac Arrhythmias by a Challenge-Best Deep Learning Neural Network）

2.2 Rhythm detection and classification

dataset(new generation of public ECG datasets)(covers a wide range of rhythms in addition to atrial fifibrillation, including different types of tachycardia, bradycardia and heart blocks):

1. the 2018 China Physiological Signal Challenge Dataset (Liu et al., 2018)（Liu, F., Liu, C., Zhao, L., Zhang, X., Wu, X., Xu, X., Liu, Y., Ma, C., Wei,S., He, Z., Li, J., & Ng, E. (2018). An Open Access Database for Evaluating the Algorithms of Electrocardiogram Rhythm and Morphology Abnormality Detection. Journal of Medical Imaging and Health Informatics, 8 , 1368–1373. doi:10.1166/jmihi.2018.2442）
2. thePTB-XL database (Wagner et al. 2020)（Wagner, P., Strodtho↵, N., Bousseljot, R., Kreiseler, D., Lunze, F., Samek,W., & Schaeter, T. (2020). PTB-XL: A Large Publicly Available electrocardiography Dataset. Scientifific Data, 7 , Article 154. doi:10.1038/s41597-020-0495-6.）
3. the Shaoxing People’s Hospital’s10,000 patients arrhythmia database (Zheng et al., 2020) ( large number of subjects and labels)（Zheng, J., Zhang, J., Danioko, S., Yao, H., Guo, H., & Rakovski, C. (2020).A 12-lead Electrocardiogram Database for Arrhythmia Research Covering More Than 10,000 Patients. Scientifific Data, 7 , Article 48. doi:10.1038/s41597-020-0386-x.）

methods:

1. standard machine learning algorithms (Mei et al., 2018)（Mei, Z., Gu, X., Chen, H., & Chen, W. (2018). Automatic Atrial Fibrillation Detection Based on Heart Rate Variability and Spectral Features. IEEE Access, 6 , 53566–53575. doi:10.1109/ACCESS.2018.2871220.）
2. employing mechanisms in deep neural networks that capture the temporal characteristic of ECG signals, such as attention and LSTM neural networks
3. attention-based deep neural network 550 (Mousavi et al., 2019)(AF detection, with F1-score of 0.994 on the MIT-BIH AF database)(for two-class classifification (AF and Non-AF rhythm).（Mousavi, S., Afghah, F., Razi, A., & Acharya, U. R. (2019). ECGNET: Learning Where to Attend for Detection of Atrial Fibrillation with Deep Visual Attention. In 2019 IEEE EMBS International Conference on Biomedical Health Informatics (BHI) (pp. 1–4). doi:10.1109/BHI.2019.8834637.）
4. convolutional neural network(Yildirim et al., 2020) (the Shaoxing People’s Hospital’s 10,000 patients arrhythmia database has been used to develop arhythm classifification model)(accuracy of 0.9224 for 7 rhythm types.)（Yildirim, O., Talo, M., Ciaccio, E. J., Tan, R. S., & Acharya, U. R. (2020). Accurate deep neural network model to detect cardiac arrhythmia on more than 10,000 individual subject ECG records. Computer Methods and Programs in Biomedicine, 197 , 105740. doi:10.1016/j.cmpb.2020.105740.）
5. Stanford Machine Learning (ML) Group (Hannun et al., 2019)（Hannun, A., Rajpurkar, P., Haghpanahi, M., Tison, G., Bourn, C., Turakhia, M., & Ng, A. (2019). Cardiologist-level arrhythmia detection and classifification in ambulatory electrocardiograms using a deep neural network. Nature Medicine, 25 , 65–69. doi:10.1038/s41591-018-0268-3.）

advantages of deep learning:

1. very little preprocessing and ECG expert knowledge is required when building the models. 

weaknesses of deep learning:

1. the resulting models are not explainable for the most part and probably for this reason, they have not been so widely used in commercial products yet