Machine learning is a widely used approach for the processing of physiological signals. This group of methods shows high performance in various problems: classification of electrocardiography (ECG) signals [4, 13], segmentation of electrocardiography signals , subject identification [6, 18, 15], prediction of a need for an urgent revascularization in emergency patients , and many others. An ECG processing also frequently complements processing other physiological signals [26, 14].
However, machine learning methods require a training dataset, and acquiring the necessary data is not an easy task. Collection of data from clinics or biological experiments with laboratory animals always requires high expenses, passing of ethical commissions, and a significant amount of time. For this reason, any open datasets for biomedical sciences are significantly smaller than datasets for other fields of a machine learning application. As an illustration, MNIST is a basic dataset for the problem of handwritten digits recognition. It contains 60,000 training images and 10,000 testing images. The biggest open dataset in PhysioNet  is the PTB Diagnostic ECG Database  that contains only 549 recordings from 294 subjects.
The small size of the training dataset raises a question about the effect of physiological variations on the performance of the proposed solutions. For example, the shape of the PQRST complex depends on the heart rate and activity of sympathetic and parasympathetic nervous systems. Thus, data acquired from the same person may strongly vary during the day. Also, a significant part of the population takes drugs that affect the T-wave position. Usually, this is QT-prolongation drugs (class III of antiarrhythmic agents) .
The major part of the studies using machine learning analysis of ECG is performed for a signal from one electrode. This may be explained by the high demand for solutions that process data recording with wearable electronics, non-invasive, and implantable systems of long-time monitoring. The human heart ECG differs between different recording leads. Thus, the performance of ECG processing with a machine learning algorithm should be dependent on the position of the surface lead.
In our current research, we are aiming to analyze the effect of human cardiac physiology phenomena on the performance of a machine learning algorithm for ECG processing. We have chosen the subject identification problem for our goal because of its simplicity and freedom in the choice of a dataset for analysis. This problem statement is similar to fingerprint identification, but it uses ECG as information about a subject that is provided by any live human body.
A solution for the problem is usually based on the methods that extract features from ECG and methods that classify the extracted features to classes, where each class is a unique subject. Also, a major part of the proposed solutions requires a short ECG fragment with length from one cardiac cycle to 5-minute records. For this reason, the subject identification problem may be solved and analyzed with almost any available dataset of electrocardiography data. This is a great advantage of this problem over the problem of ECG classification, PQRST complex segmentation, and many others. A wide description of subject identification methods may be found in recent reviews[6, 18].
In this preliminary study, we used a solution that was proposed in the previous work of the co-authors [2, 1]. That solution used the morphology of PQRST complexes and indirectly included information about the heart rate variability. We test this solution against the choice of ECG lead position, long-time variability of heart rate, and the effects of QT-prolongation drugs (class III). In the discussion, we compare our results with other observations and show how our findings may be helpful to wide area of studies that use machine learning approaches to ECG analysis.
Physikalisch-Technische Bundesanstalt Diagnostic ECG Database (PTB Database) [3, 7] was chosen for analysis of the effect of a lead choice on the subject identification problem. This database includes 549 records from 290 subjects. Each subject is represented by one to five records. Each record includes the conventional 12 leads (I, II, III, aVR, aVL, avf, V1, V2, V3, V4, V5, V6), and the three Frank leads (Vx, Vy, Vz).
The Long Term ST Database (LTSTD) [11, 7] was chosen for analysis of the effect of 24-hours rhythm variations to the subject identification problem. The LTSTD contains 86 lengthy ECG recordings of 80 human subjects, chosen to exhibit a variety of events of ST-segment changes, including ischemic ST episodes, axis-related non-ischemic ST episodes, episodes of slow ST level drift, and episodes containing mixtures of these phenomena.
In addition to ischemic events, any long-term ECG recording includes slight variations of the PQRST complex caused by the regulation of parasympathetic and sympathetic nervous systems, which is related to a person’s diurnal cycles and some stress surroundings.
The ECG Effects of Ranolazine, Dofetilide, Verapamil, and Quinidine database (ECGRDVQ database) [12, 7] was chosen for analysis of QT-prolonging drugs on the person identification problem. That database contains ECG recordings of 22 healthy subjects for 24 hours under the effect of dofetilide (500 μg), quinidine sulfate (400 mg), ranolazine (1500 mg), verapamil hydrochloride (120 mg). QT-prolonging drugs affect on duration of the transmembrane action potential of cardiomyocytes, which causes changes in the T-wave shape and QT interval prolongation.
|naive Bayes classifier||54%||45%||46%||41%||46%||45%||53%||60%||60%||58%||52%||49%||41%||19%|
|decision tree classifier||69%||70%||67%||69%||70%||69%||72%||73%||76%||76%||70%||74%||69%||9%|
|k-nearest neighbour votes||92%||91%||89%||86%||93%||89%||93%||95%||93%||94%||92%||95%||86%||9%|
|linear discriminant analysis||92%||91%||82%||82%||89%||87%||86%||87%||86%||86%||88%||88%||82%||10%|
linear support vector classifier
|logistic regression classifier||95%||97%||94%||92%||96%||95%||96%||97%||97%||98%||96%||96%||92%||6%|
|nearest centroid classifier||62%||53%||51%||50%||53%||53%||57%||63%||61%||61%||56%||59%||50%||13%|
|random forest classifier||10%||10%||8%||13%||14%||7%||9%||7%||6%||7%||10%||10%||6%||8%|
|ridge regression classifier||39%||46%||37%||36%||44%||39%||38%||43%||43%||39%||38%||38%||36%||10%|
|Gaussian mixture model||63%||52%||53%||52%||57%||54%||66%||73%||72%||67%||60%||58%||52%||21%|
|support vector machine||70%||66%||57%||63%||62%||59%||65%||66%||67%||67%||65%||64%||59%||13%|
|naive Bayes classifier||79%||63%||16%||90%|
|decision tree classifier||91%||52%||39%||93%|
|k-nearest neighbour votes||97%||90%||7%||96%|
|linear discriminant analysis||95%||90%||5%||97%|
|linear support vector classifier||90%||80%||10%||97%|
|logistic regression classifier||96%||80%||16%||98%|
|nearest centroid classifier||81%||83%||-2%||89%|
|random forest classifier||47%||41%||6%||67%|
|ridge regression classifier||72%||69%||3%||74%|
|Gaussian mixture model||95%||90%||5%||97%|
|support vector machine||70%||90%||-20%||95%|
Ii-B ECG processing
In the current study, the subject identification problem is considered as a classification problem that should be solved with machine learning approaches. Each ECG signal is represented as a vector that was used as input of a classification algorithm, and each subject (individual) corresponds to the target class of the algorithm output. In this case, the identification ratio is equivalent to classification accuracy. The subject identification problem was established for the short ECG recorded from a single lead with a length of 20 heartbeats.
The processing procedure was as follows. First, the whole signal was separated in individual heartbeat signals. Then, the heartbeat signals were aligned to the peak of the R wave. The position of each peak in time and its amplitude provided nine features for a heartbeat. Features that were extracted from all heartbeats of the short fragment were joined together for the creation of a long input vector of features with 180 components.
The performance analysis was performed for each database independently. The training dataset was formed from the vectors corresponding to the first 20 heartbeats of the first record of each patient. Validation datasets were formed in various ways. For the PTB database, the only one validation dataset was formed from randomly chosen ECG fragments of the normal rhythm for each patient. For the LTSTB and ECGRDVQ databases, several validation datasets were formed for each half-hour of recordings. These datasets include the first 20 complexes at the beginning of each half-hour intervals of long-time monitored ECG.
The 14 algorithms from the scikit-learn package 
were used for classification: a multi-layer perceptron, naive Bayes classifier for multivariate Bernoulli distributions, a decision tree classifier, an extra-trees classifier, k-nearest neighbor votes, a linear discriminant analysis, a linear support vector classifier, a logistic regression classifier, a nearest centroid classifier, a random forest classifier, a ridge regression classifier, a ridge classifier with built-in cross-validation, a Gaussian mixture model, and a support vector machine. The extended description and analysis of the subject identification problem also can be found in the previous work of the co-authors[2, 1].
Table I shows the analysis of subject identification accuracy for each of the 12 conventional ECG leads. The presented table shows that methods with high accuracy are almost independent of the lead choice for subject identification. This proposition is confirmed with non-parametric rank correlation coefficients: corr.=-0.52 Spearman’s R (), corr.=-0.411 Kendall’s (0.046). The best accuracy was obtained using three methods: multi-layer perception, extremely randomized tree classifier, and logistic regression. These methods also provide the smallest difference between the minimal and maximal accuracy values across all leads. Thus, we should assume that the performance of machine learning approaches is almost independent of chosen ECG leads.
Figure 1 presents an analysis of subject identification accuracy under normal 24-hour rhythm variability. The series of validation datasets is built-on short ECG fragments that are spaced through 30-minute intervals on the timeline. The minimal variations in accuracy observed for the multi-layer perceptron, extremely randomized tree, and support vector machines.
Accordingly, the structure of plots, we conclude that algorithms with high accuracy are almost independent of the variability of ECG during the 24-hour, long-time monitoring period.
Table II shows the effect of QT-prolongation drugs on the problem of subject identification by ECG. This effect cannot be ignored in a practical application because accuracy reduction may reach 40%. However, the accuracy of the algorithm is recovered if the training dataset is extended with ECG signals after taking of the drug.
We observe weak dependency between the performance of the algorithm on normal conditions and performance of the algorithm after taking of drugs if this algorithm does not have data with prolonged QT interval in the training dataset (corr.=0.52, , Spearman’s R; corr.=0.40, p¡0.05m Kendall’s ). However, machine learning approaches with high performance on normal data, show high performance on both types of data if the training dataset is extended with additional cases (corr.=0.80, , Spearman’s R; corr.=0.64, , Kendall’s ).
show a significant difference in morphological parameters of heartbeats in ECG signals from different leads. All standard leads show various signal-to-noise ratio and distortion relating to body movements and electrical muscle activity. Also, the chest ECG lead is necessary for the detection of regions of myocardial infarction and cardiac ischemia.
On the other hand, a lot of problems for ECG analysis may be solved via single-lead ECG. Many works from the PhysioNet challenge  and early works  show the possibility to classify the atrial fibrillation and differentiate the supra-ventricular tachycardia from ventricular tachycardia. Recent work  classifies 12 heart rhythm disturbances from only one lead with cardiologist level performance.
Review  reports about the possibility to solve the subject identification problem for one lead, and that opinion is supported by at least ten works that have shown this possibility before.
In the present study, we analyzed the effect of the ECG lead choice on the subject identification problem. Our study shows that algorithm performance is almost independent of lead choice. This observation is consistent with the results of prior research .
The negative effect of the intra-subject ECG variation in time on subject-identification rate is well described in the literature. Especially, a significant reduction in the identification ratio has been observed in the comparison between recordings in rest and exercise conditions.
Work  reported a significant reduction of subject identification rate from 96% to 89% (7% difference) in 120 minute monitoring.
Work  compared algorithm performance for ECG recorded for the subject in a supine and standing positions, and while doing exercise on a bicycle. This work shows a variety of subject identification rates of up to 8.1% between different physiological conditions for the algorithm with the best prediction rate. Also, the algorithm shows increasing a subject prediction rate when the training dataset is enriched with data from all available physiological conditions.
Work  showed similar findings. This work tested four different approaches to data acquired from subjects in five different postures: sitting, standing, tripod, supine, and exercise. The study used alternative metrics for algorithm performance evaluation and cannot be directly compared with our results. However, the obtained results show that changing physiological conditions reduce the accuracy of any algorithm.
Intra-individual variation of ECG rhythm is also mentioned or was the main motivation in several other works [28, 23, 24]. Also, changes in ECG rhythm between the rest and exercise conditions are the basis of ECG stress testing that is a widespread method in clinical diagnostics. However, we could not find any work that studied the effect of ECG variation and human exercise load on the ECG classification problem or other types of problem statements for ECG processing.
Our observation also shows a reduction in subject identification ratio (classification accuracy) due to ECG changes over time. However, some of the classification algorithms in our study show a reduction in performance that is not more than 5%. This result is smaller than presented in other works [20, 19]. This could be explained by the more effective approach taken here, and the absence of physical stress in patients who provide long-time monitored ECG for LTSTD.
We observe significant negative effects of QT-prolongation drugs to subject-identification ratio. Compensation of the algorithm performance decreasing requires enrichment of training dataset with ECG signals recorded under the effect of QT-prolongation agents. Effect of drugs on the subject identification problem is not mentioned in reviews of 2015  and 2019  years. We try to find similar observations for other problems of ECG processing. One of such observation is mentioned in . Effect of sotalol (QT-prolongation drugs) reduces the dispersion of QT interval for the patient with Long-QT syndrome. This is notable because QT dispersion is the main diagnostic criteria of this genetic disease. In other words, the treatment of the disease decreases the ability to detect this disease.
Several generalizations can be drawn from the subject identification problem on other methods of ECG processing with machine learning approaches. We can speculate that a wide area of problem statements for ECG processing may be solved with only a signal from one lead if this problem is not related to very small changes of ECG (some myocardial infections, myocardial ischemia). ECG variability in 24-hours and ECG changes in stress should lead to a decrease in the performance of an algorithm that processes short ECG fragments (0.5-30 min). This becomes especially important if an algorithm is designed to process data from subjects with high emotional and physical stress, but a dataset for the training is acquired in the same period of a day, and all patients stay in the rest condition. The effect of the QT-prolongation drug may be crucial for the performance of any algorithm that analyzes the properties of normal rhythm.
Issues with ECG variability in 24-hours, ECG changes in stress conditions, and QT-prolongation agents may be solved with enrichment of the training dataset with ECG from a subject exposed to the conditions that affect cardiac electrophysiology. This recommendation is based on our observation with the ECGRDVQ database and earlier reports [19, 27].
In this study, we have analyzed variations in the performance of machine learning approaches for ECG processing depending on the effects of human physiology on the input ECG signals. We evaluated effects of variations in ECG recordings caused by different positions of ECG leads, a variation in the heart rhythm during 24-hours, and the effects of QT-prolongation drugs. Our analysis was performed for the subject identification problem, but we expect that our observations may be widespread onto other problems of ECG analysis with machine learning approaches.
We found that the solution of the subject identification problem is significantly sensitive to ECG changes caused by taking QT-prolongation agents. Also, ECG variability in 24-hours negatively affects the performance of the algorithms. Other publications show that this effect may be stronger in the case of physical load.
Based on our results and literature observations, we suppose that any machine learning approach to ECG processing is sensitive to three physiological phenomena: ECG variability in 24-hours, ECG changes in the stress condition, and QT segment prolongation due to effects of class III antiarrhythmic agents. The reduction in the machine learning performance in real applications may be overcome by the enrichment of the training dataset with ECG acquired when the subjects are exposed to physical load and/or take medications (particularly, QT-prolongation agents) during ECG evaluation.
-  (2018) Statistical assessment of biometrical signs by electrocardiography. Russian Journal of Cardiology 23 (7), pp. 84–91. External Links: Cited by: §I, §II-B.
-  (2018) Factors influencing accuracy of biometrical personal identification based on cardiograms. Pattern Recognition and Image Analysis 28 (3), pp. 421–426. Cited by: §I, §II-B.
-  (1995) Nutzung der ekg-signaldatenbank cardiodat der ptb über das internet. Biomedizinische Technik/Biomedical Engineering 40 (s1), pp. 317–318. Cited by: §I, §II-A.
-  (2017) AF classification from a short single lead ecg recording: the physionet/computing in cardiology challenge 2017. In 2017 Computing in Cardiology (CinC), pp. 1–4. Cited by: §I, §IV.
-  (1990) QT dispersion: an indication of arrhythmia risk in patients with long qt intervals.. Heart 63 (6), pp. 342–344. Cited by: §IV.
-  (2015) Individual identification via electrocardiogram analysis. Biomedical engineering online 14 (1), pp. 78. Cited by: §I, §I, §IV, §IV, §IV.
-  (2000) PhysioBank, physiotoolkit, and physionet: components of a new research resource for complex physiologic signals. Circulation 101 (23), pp. e215–e220. Cited by: §I, §II-A, §II-A, §II-A.
-  (2019) Artificial intelligence to predict needs for urgent revascularization from 12-leads electrocardiography in emergency patients. PloS one 14 (1), pp. e0210103. Cited by: §I.
Cardiologist-level arrhythmia detection and classification in ambulatory electrocardiograms using a deep neural network. Nature medicine 25 (1), pp. 65. Cited by: §IV.
-  (2001) Geometrical aspects of the interindividual variability of multilead ecg recordings. IEEE Transactions on Biomedical Engineering 48 (5), pp. 551–559. Cited by: §IV.
-  (2003) Long-term st database: a reference for the development and evaluation of automated ischaemia detectors and for the study of the dynamics of myocardial ischaemia. Medical and Biological Engineering and Computing 41 (2), pp. 172–182. Cited by: §II-A.
-  (2014) Differentiating drug-induced multichannel block on the electrocardiogram: randomized study of dofetilide, quinidine, ranolazine, and verapamil. Clinical Pharmacology & Therapeutics 96 (5), pp. 549–558. Cited by: §II-A.
Real-time patient-specific ecg classification by 1-d convolutional neural networks. IEEE Transactions on Biomedical Engineering 63 (3), pp. 664–675. Cited by: §I, §IV.
Classification of the physical training level by heart rate variability and stabilography data.
2017 Siberian Symposium on Data Science and Engineering (SSDSE), pp. 49–54. Cited by: §I.
-  (2005) Biometric human identification based on electrocardiogram.[master’s thesis] electrotechnical university “leti”. Saint-Petersburg, Russian Federation. Cited by: §I.
-  (1999) Evolution, mechanisms, and classification of antiarrhythmic drugs: focus on class iii actions. The American journal of cardiology 84 (9), pp. 11–19. Cited by: §I.
-  (2011) Scikit-learn: machine learning in python. Journal of machine learning research 12 (Oct), pp. 2825–2830. Cited by: §II-B.
-  (2018) Evolution, current challenges, and future possibilities in ecg biometrics. IEEE Access 6, pp. 34746–34776. Cited by: §I, §I, §IV.
-  (2016) ECG biometric analysis in different physiological recording conditions. Signal, image and video processing 10 (2), pp. 267–276. Cited by: §IV, §IV, §IV.
-  (2013) Influence of qt correction on temporal and amplitude features for human identification via ecg. In 2013 IEEE Workshop on Biometric Measurements and Systems for Security and Medical Applications, pp. 22–27. Cited by: §IV, §IV.
-  (2000) Intra-individual variability of the electrocardiogram. Assessment and exploitation in computerized ECG analysis [Ph. D. thesis]. Cited by: §IV.
-  (2019) ECG segmentation by neural networks: errors and correction. In 2019 International Joint Conference on Neural Networks (IJCNN), pp. 1–7. Cited by: §I.
-  (2014) ECG biometric with abnormal cardiac conditions in remote monitoring system. IEEE Transactions on systems, man, and cybernetics: systems 44 (11), pp. 1498–1509. Cited by: §IV.
-  (2011) Automobile driver recognition under different physiological conditions using the electrocardiogram. In 2011 Computing in Cardiology, pp. 753–756. Cited by: §IV.
-  (2012) Third universal definition of myocardial infarction. Circulation 126 (16), pp. 2020–2035. Cited by: §IV.
-  (2019) Case study of interrelation between brain-computer interface based multimodal metric and heart rate variability. In 12th International Conference on Health Informatics, HEALTHINF 2019-Part of 12th International Joint Conference on Biomedical Engineering Systems and Technologies, BIOSTEC 2019, pp. 532–538. Cited by: §I.
-  (2014) On evaluating ecg biometric systems: session-dependence and body posture. IEEE Transactions on Information Forensics and Security 9 (11), pp. 2002–2013. Cited by: §IV, §IV.
-  (2014) Research on ecg biometric in cardiac irregularity conditions. In 2014 International Conference on Medical Biometrics, pp. 157–163. Cited by: §IV.