Detection of Subtle ECG Changes Despite Superimposed Artifacts by Different Machine Learning Algorithms

Abstract

Analyzing electrocardiographic (ECG) signals is crucial for evaluating heart function and diagnosing cardiac pathology. Traditional methods for detecting ECG changes often rely on offline analysis or subjective visual inspection, which may overlook subtle variations, particularly in the case of artifacts. In this theoretical, proof-of-concept study, we investigated the potential of five different machine learning algorithms [random forests (RFs), gradient boosting methods (GBMs), deep neural networks (DNNs), an ensemble learning technique, as well as logistic regression] to detect subtle changes in the morphology of synthetically generated ECG beats despite artifacts. Following the generation of a synthetic ECG beat using the standardized McSharry algorithm, the baseline ECG signal was modified by changing the amplitude of different ECG components by 0.01–0.06 mV. In addition, a Gaussian jitter of 0.1–0.3 mV was overlaid to simulate artifacts. Five different machine learning algorithms were then applied to detect differences between the modified ECG beats. The highest discriminatory potency, as assessed by the discriminatory accuracy, was achieved by RFs and GBMs (accuracy of up to 1.0), whereas the least accurate results were obtained by logistic regression (accuracy approximately 10% less). In a second step, a feature importance algorithm (Boruta) was used to discriminate which signal parts were responsible for difference detection. For all comparisons, only signal components that had been modified in advance were used for discretion, demonstrating that the RF model focused on the appropriate signal elements. Our findings highlight the potential of RFs and GBMs as valuable tools for detecting subtle ECG changes despite artifacts, with implications for enhancing clinical diagnosis and monitoring. Further studies are needed to validate our findings with clinical data.

1. Introduction

Electrocardiographic (ECG) signal analysis allows the identification of heart function and integrity [1]. The offline identification of distinct ECG changes as indicators of different pathologies has been one of the cornerstones of both cardiovascular and non-cardiovascular medicine for decades [2,3]. However, besides the traditional 12-lead ECG interpretation in clinical cardiology, there are numerous other clinical situations where ECG changes are continuously monitored [4]. For example, the use of wearable devices for cardiovascular monitoring including ECG has gained significant interest and clinical value for patients and healthcare workers [5]. Despite recent technological advances, certain limitations continue to compromise the accuracy of signal interpretation, potentially leading to imprecise pathology detection. These lead to missing variances of the ECG morphology due to multiple confounders such as internal or external artifacts, the low resolution of monitors used, the impossibility of analyzing changes of the signal obtained on the fly, and the general problem of inaccuracies of ECG traces and intervals [6]. Computer-aided algorithms might serve as promising tools to overcome these shortcomings. However, most of the technical solutions used so far in clinical practice suffer from low sensitivities in detecting subtle changes, particularly in noisy ECG signals [7,8,9,10].

Modern machine learning tools have the theoretical potential to detect differences between two ECG signals despite the aforementioned limitations. Although they are all based on different mathematical approaches, they represent different supervised entities with the same underlying principle. The main idea of this approach is to train a model on well-known data and evaluate it using unseen examples, where the correct outcome is known [11,12,13,14].

In this study, we aimed to determine the diagnostic accuracy of five different machine learning algorithms to detect subtle changes in the morphology of synthetically generated ECG signals despite superimposition by different degrees of signal noise. Our hypothesis was that machine learning algorithms could reliably detect even subtle changes in noisy ECG signals.

2. Materials and Methods

The investigation was designed as a theoretical proof-of-concept study conducted at the Department of Anaesthesiology and Critical Care Medicine at the Kepler University Hospital, a tertiary center in Linz/Austria. Since only synthetical data were used, no ethical approvement was required.

To evaluate the ability of five machine learning methods to correctly classify changes in the morphology of noisy ECG beats, the following steps were performed. (1) A synthetical ECG signal was generated. (2) This ECG signal was then modified by changing the amplitudes of different ECG components. (3) The different synthetic ECG signals were then superimposed with variable degrees of artifacts. (4) Five machine learning algorithms were then applied to detect changes in the ECG signal when compared to the originally generated synthetic ECG trace. The discriminatory accuracy was used to report the predictive value of each machine learning algorithm. Finally, the Boruta algorithm was used to determine which components of the modified ECG signal were important to detect changes between the modified and baseline ECG trace.

2.1. Synthetic ECG Generation and Modification

The normal ECG consists of a typical pattern of waves and segments. The initial small, positive deflection is called the P wave, which represents atrial depolarization, typically lasting no more than 0.1–0.11 s. Deviations in presence, shape, and duration of the P wave might be indicators of pathologies of either the sinus node, the primary intrinsic cardiac pacemaker, or the atrial conduction system. Following the P wave, the ECG trace then returns to the isoelectric baseline (PR segment) before continuing with a short and small negative deflection (Q wave), followed by a sharp positive deflection (R wave), eventually followed by a final negative deflection (S wave). This so-called QRS complex reflects the depolarization of the ventricles. Duration of the QRS complex is defined as the time interval between the beginning of the Q wave and the end of the S wave, and its duration under normal and physiologic conditions is <0.12 s. A prolonged or deformed QRS complex is an indicator of pathologies of the intraventricular electrical conduction system due to ischemia, inflammation, or degenerative processes. Immediately after ventricular depolarization, the ECG trace remains isoelectric (ST segment), before changing its vector to a final positive deflection, the T wave, representing fast ventricular repolarization. Deviations in shape, height, or vector of deflection are indicators of different cardiac pathologies (e.g., negative T waves in myocardial ischemia or positive peaked T waves in hyperkalemia).

The standardized McSharry algorithm was used to generate a realistic synthetic ECG trace [15]. This algorithm reflects a data-driven approach to modeling the heart’s electrical activity by generating a trajectory in a three-dimensional space. Distinct points of the ECG trace, such as the P, Q, R, S, and T waves, are described by events corresponding to negative and positive attractors/repellers in the direction [15]. Using this algorithm, one single ECG beat with a sampling frequency of 256 Hz was generated as a baseline for all further analyses.

Based on the original publication by McSharry et al., the dynamical equations of motion are defined by three different equations [15]:

$$\begin{array}{cc}\hfill \dot{x}& =\alpha x-\omega y\hfill \\ \hfill \dot{y}& =\alpha y+\omega x\hfill \\ \hfill \dot{z}& =-{\displaystyle \sum _{i\in \{P,Q,R,S,T\}}}{a}_i\Delta {\theta }_i\mathrm{e}\mathrm{x}\mathrm{p}\left(-{\displaystyle \frac{\Delta {\theta }_i^2}{2b_i^2}}\right)-\left(z-z_0\right)\hfill \end{array}$$

where $\alpha =1-\sqrt{x^2+y^2}, \Delta {\theta }_i=\left(\theta -{\theta }_i\right)\mathrm{m}\mathrm{o}\mathrm{d}2\pi , \theta =\mathrm{a}\mathrm{t}\mathrm{a}\mathrm{n}2(y,x)$ (the four-quadrant arctangent of the real parts of the elements of $x$ and $y,$ with $-\pi \le \mathrm{atan}2\left(y,x\right)\le \pi ),$ and $\omega$, as its moving around the limit cycle, is defined as the angular velocity of trajectory.

Additionally, as mentioned by McSharry et al. [15], a numerical integration was added to the abovementioned algorithm using a fourth-order Runge–Kutta method at a fixed time step [16].

In a second step, further ECG beats were generated with an increase in the P, Q, R, S, and T wave amplitudes by 0.01, 0.02, 0.03, 0.04, 0.05, and 0.06 mV, respectively. The magnitude of these changes is so small that none can be detected easily by a cursory visual analysis of the signal, especially in noisy signals. Subsequently, to simulate the typical noise of a clinical ECG trace, each of these 30 distinctive ECG beats (6 different amplitudes for each wave) was duplicated 100 times with an identical jitter (0.1 mV, 0.2 mV, 0.3 mV, respectively), resulting in three sets of 100 beats of identical features but different individual shapes for each of the 30 distinctive ECG signals. Finally, after variation of relevant features of the ECG beats and deconstruction of the initial synthetic ECG, each of the 30 groups of similar ECG beats underwent classification against the beats of the baseline ECG trace. In total, 3100 ECG beats were analyzed in this way.

2.2. Machine Learning Algorithms Used

Random forests (RFs), gradient boosting machines (GBMs), deep neural networks (DNNs), an ensemble learning technique, as well as logistic regressions as the standard comparator were applied to detect differences between the modified ECG beats and the baseline ECG trace.

Random forests (RFs) are a type of supervised machine learning algorithm for both classification and regression tasks first proposed by Leo Breiman [11]. The main principle of RFs is to construct several different, uncorrelated decision trees during the training phase and subsequently merge them together to create a single output. A random subset of features from a random subset of the dataset is generated using feature randomness, thereby aiding in constructing uncorrelated forests. Additionally, feature bagging or bootstrap aggregation can increase randomness and diversity of the dataset [17]. Advantages are its reduced risk of over-fitting, the ability to handle large datasets, the ability to provide both classification and regression tasks, as well as its ability to determine feature importances. However, RFs lack the ability to obtain an interpretation of causal links between predictors and response, thereby limiting their use in certain research fields [18].

Gradient boosting machines (GBMs) are an ensemble machine learning method derived by Jerome Friedman [19]. GBMs can be considered as an ensemble model of different, sequentially trained decision trees. The principle behind this algorithm is to consecutively fit new models to provide a more accurate estimate of the response variable and to construct new base-learners to be highly correlated with the negative gradient of the loss function, associated with the whole ensemble [20]. Nevertheless, GBMs are prone to over-fitting, require additional pre-processing of the data to ensure adequate model performance, require balanced datasets for classification tasks, and are associated with higher resource consumption, since extended training time for the training models is required.

Deep neural networks (DNNs) are a type of deep learning (DL) machine learning method. DNNs are fully connected artificial neural networks (ANNs) consisting of input and output layers as well as multiple hidden layers in between. The input layer receives input data whilst the output layer performs the decision or prediction process regarding the input signal. The hidden layers of the artificial neural network are the defining part of the neural network architecture and allow the transformation of inputs, the approximation of functions, or the capturing of data patterns [21,22]. DNNs can be supervised, unsupervised, or a combination of both. Deep learning methods can process vast amounts of features and are therefore useful when dealing with a large amount of data. They have been successfully integrated or could be practically applied in the future in several different fields ranging from robotics to natural language processing, from cybersecurity to speech recognition [21]. Nevertheless, due to both their weak statistical interpretability and poor reasoning regarding the achievement of their results and decisions, DL and DNNs are often considered “black-box” solutions.

Ensemble learning methods are an alternative to deep learning methods. Ensemble learning refers to a concept that combines different machine learning models. Different baseline models (“weak learners”) are trained to solve the same problem by using ensemble methods such as bagging, stacking, boosting, or averaging. Combining several individual models to improve prediction performance and outperformance of other machine learning models is the main advantage of ensemble learning methods [23].

Logistic regression analysis is a statistical model that is used to test the association of both categorical and continuous independent variables with a single dichotomous dependent variable [24]. Logistic regression analysis is based on a sigmoid function. It is a commonly used statistical model, easy to apply, with high accuracies and only a small risk of over-fitting. One of its major limitations is the assumption of linearity between the dependent and independent variables. We chose logistic regression as the baseline comparator for the other machine learning algorithms.

Since we designed a balanced experimental approach investigating a symmetrical problem, we chose the discriminatory accuracy as the main outcome variable of our study to compare the ability of the abovementioned machine learning algorithms to detect subtle ECG changes despite superimposed artifacts. Accuracy was calculated according to the following formula:

Accuracy = (TP + TN)/(TP + TN + FN + FP)

where TP represents the true positives, TN represents the true negatives, FN represents the false negatives, and FP represents the false positives.

This appears to be in line with daily clinical practice; the higher the number of correct classifications, the better the respective algorithm. In view of the numerous entities that were generated (different changes in amplitude of different ECG components all superimposed by three degrees of jitter), we refrained from applying statistical tests to compare the discriminatory accuracy between the five machine learning models. The superiority of different algorithms was judged by overall “eyeballing” the differences over the resulting heat maps depending on the algorithms used.

All calculations were performed using standardized statistical software (R^® version 4.1.2; R^® Foundation for Statistical Computing, Vienna, Austria) and the H₂O package (R^® Interface for H₂O; R^® package version 3.42.0.2). The S wave in the QRS complex of each ECG sweep was automatically classified by wavelet analyses, and data were cut to enable the data snippets for each heartbeat to start precisely 350 ms before the R wave and last until 520 ms after the R wave. As a quality control measure, data snippets were visually controlled for the presence of calculation artifacts. The latencies of the defined ECG time points (P, Q, R, S and the end of the T wave) as well as the maximum amplitude of the T wave were preset by parameters of the McSharry algorithm. For our random forest and gradient boosting methods, we used 100 decision trees with a maximum depth of 5 as base learners to balance performance and computing time. For the DNN model, we constructed a 50-hidden-layer network with 100 neurons per layer and applied the Rectified Linear Unit activation function and the Adam optimizer.

2.3. Feature Importance

The RF-based Boruta algorithm [25] was used to determine which components of the modified ECG signals were most important in detecting differences between the modified and baseline ECG traces. The Boruta algorithm generates randomly permuted copies of all variables as shadow attributes and feeds them into RFs to generate importance (Z-scores) for all attributes. The algorithm accepts variables with an importance significantly higher than the maximal importance of random attributes. Thus, it is possible to determine which time points are the most important for the algorithm to classify the ECG signal.

2.4. Proof-of-Concept Control Measurements

In a further step, we created two synthetic ECG signals of common cardiac pathologies: ST segment depressions indicating myocardial ischemia on the one hand, and Long-QT syndrome (LQTS), a type of hereditary or acquired heart rhythm disorder and a relevant predisposing risk factor for arrythmias. Again, a Gaussian jitter of 0.1–0.3 mV was overlaid to simulate artifacts. As performed with the initial physiologic ECG trace, the RF-based Boruta algorithm [25] was applied in the same way as described above, to determine which components of the modified ECGs were most important for detecting differences between the modified and the baseline ECG. By applying these control measurements, we tried to exclude the potential bias posed by the algorithm learning features other than the ones we initially modified, and aimed to demonstrate that the actual shape of the ECG only plays a minor role regarding the ability of the individual machine learning algorithms to detect subtle ECG changes despite artifacts.

3. Results

The synthetically generated physiologic baseline ECG trace is shown in Figure 1.

When the amplitude change was in the range of 5–10% of the superimposed jitter (0.01 mV–0.02 mV), all models failed to classify the different beats with an acceptable accuracy (Figure 2).

However, with an increase in the differences in the amplitude of the respective ECG components of 0.04 mV and more, the accuracies of RF, GBM, and ensemble learning methods in detecting differences in the ECG pattern compared to the baseline ECG increased and reached discriminatory accuracies of ≥0.8 for all components of the ECG beat. However, the ability of all machine learning algorithms to detect changes in the modified ECG trace compared to the baseline ECG progressively decreased when sequentially increasing the identical Gaussian jitter from 0.1 mV to 0.3 mV.

Nevertheless, the accuracy of RF, GBM, and ensemble learning methods in detecting occurring differences in the P wave (reflecting depolarization of the atria) and the T wave (reflecting repolarization of the ventricles) remained at a constant level, ranging from 0.8 to ≥0.9 irrespective of the Gaussian jitter applied at an amplitude difference of 0.04 mV or more (Figure 2).

The Boruta algorithm indicated that the features with the highest influence were the components of the ECG trace where the amplitude was artificially increased (Figure 3). No other parts of the ECG signal played a relevant role.

When investigating the potential of the different machine learning algorithms in a pathologic control ECG trace depicting ST segment depressions, we found that both RFs and GBMs demonstrated excellent accuracies of ≥0.98 in detecting changes in the ST segment (1.5 mm versus 2.0 mm ST segment depression) when a Gaussian jitter of 0.1 mV was superimposed. When sequentially increasing the identical jitter to 0.3 mV, the discriminatory power of both RFs and GBMs decreased, but remained high, with stable accuracies >0.96 (Figure 4A). Again, the Boruta algorithm indicated that the features with the highest influence were the components of the ECG trace where the amplitude was artificially modified (Figure 5). No other parts of the ECG besides the ST segment signal played a relevant role. Similar results were found when investigating the ECG trace of Long-QT syndrome (LQTS). Again, both RFs and GBMs demonstrated their discriminatory power for detecting a prolongation of QT interval (442 ms versus 450 ms) in the synthetic ECG trace, reaching accuracies of 1.0 and ≥0.96, at a Gaussian jitter of 0.1 mV and 0.2 mV, respectively (Figure 4B). The relevant comparison of feature importances in ECG pattern differentiation using the Boruta feature selection algorithm is displayed in Figure 5.

Interestingly, in the control ECG traces depicting ST segment depressions and LQTS, the ensemble learning method performed better than in the initial physiologic ECG trace, demonstrating high discriminatory power with excellent accuracies for difference detection (Figure 4.)

4. Discussion

In this theoretical, proof-of-concept study, we report the discriminatory accuracies of five different machine learning algorithms in detecting subtle changes in the morphology of synthetically generated ECG signals superimposed with different degrees of artifacts. Using this experimental approach, we found that machine learning algorithms were highly discriminative for classifying the amplitude of specific parts of the physiologic ECG, especially the P and T waves, even if the signal-to-noise ratio was in the range of 0.1. The best classification results were obtained by both RFs and GBMs, whereas logistic regression yielded the lowest discriminative power.

To the best of our knowledge, this is the first study to demonstrate the feasibility of detecting minimal or imperceptible changes of single ECG traces superimposed by artifacts with the help of machine learning algorithms. Our approach differs from the usual data interpretation performed so far; instead of using a longer ECG registration, we identified single beats and aligned them by the R wave. Using this method, it was possible to interpret a time point of the ECG signal relative to the R wave as a feature for the machine learning algorithm without analyzing the signal by a convolution layer of a neural network.

Machine learning tools, particularly RFs, GBMs, DNNs, and ensemble learning methods, are attracting more and more interest as a solution for problems in supervised learning algorithms in the medical field. During the last few years, these algorithms have become the method of choice for several different settings in supervised pattern recognition in datasets. Typically, they are used to train classification models, allowing separate samples of different classes based on different features and estimating which features were important for this classification [26]. For example, the RF algorithm has become very popular for providing two aspects important for data mining: (1) high prediction accuracy and (2) information on variable importance for classification. The prediction performance of RFs compares well to other classification algorithms such as support vector machines or artificial neural networks. The five algorithms that we have chosen for our study are known to work well in medical tabulated data. Our selection of potential algorithms that could be applied for the detection of subtle ECG changes is far from complete. Other machine learning tools have been proposed as well for medical data and could have been used for our data analysis as well [27]. For example, 1D convolutional neural networks, a feed-forward artificial neural network consisting of alternating convolutional and subsampling layers [28], may theoretically have resulted in higher and more consistent accuracies in detecting changes in ECG morphology. However, for this approach, a large amount of training data is necessary, which can clearly be avoided by our approach. Additionally, since two of our tested algorithms resulted in very good results regarding the discriminatory accuracy, we refrained from testing more potential algorithms, since we expected that the results might be similar. However, we therefore cannot rule out the idea that another machine learning algorithm might be even better at classifying ECGs.

The results of our study imply certain possibilities for future application in clinical medicine. For example, our results demonstrate that RFs and GBMs can reliably differentiate ECG morphologies even when substantial artifacts are present. Despite the fact that it is well known that RFs and GBMs often work best for classification tasks of tabular, medical data, the exact and precise reasons why some machine learning methods and algorithms work better at medical classification tasks than others are very difficult to define [29]. We therefore cannot really explain the reasons why RFs and GBMs performed that well in our study.

Nevertheless, our approach might be promising and could be used to compare real-time to baseline ECG traces in different physiologic situations and settings, even when relevant noise is present in the ECG signal. The theoretical applicability is wide, ranging from integration into wearable cardiovascular monitoring devices to continuous ECG monitors in healthcare settings. Thereby, machine learning algorithms detecting pre-defined ECG trace changes could enable targeted screening for respective ECG changes, thus identifying distinct pathologies (e.g., increases in the amplitude of the T wave, as seen in hyperkalemia). As only changes in ECG signals that have been defined a priori can be detected when using this approach, using large databases which include physiological and expert-diagnosed, pathologic ECG traces could inform appropriately trained models to allow for early detection of ECG pathologies equivalent to the accuracy of an experienced examiner.

Several limitations need to be considered when interpreting the results of our study. First, we used synthetically generated ECG signals instead of ECG traces derived from real patients. This limits the generalizability of our findings to the clinical setting. Furthermore, when using our approach with clinical ECG data, pre-processing of the ECG trace must be performed by removing a baseline wander and applying different filters to account for potential artifacts caused by respiration or body movements that could lead to ECG beat morphology changes without cardiac origin [30]. Baseline wander noise frequencies are typically in the range of 0–0.5 Hz [31]. Using pre-processing, baseline noise can be suppressed while clinically important information remains visible. It remains unknown if such an approach would have influenced our results. Next, we generated our ECG sweeps beat by beat. As a consequence, the main shape of every sweep does not change over time. Therefore, in a real-life scenario, the classification tasks we performed in our simplified model might be more difficult in a clinical setting. However, since we did not use real clinical data, we cannot guess how much this influences our results. Furthermore, our study did not investigate whether all the confounders associated with conventional ECG trace monitoring (e.g., internal and external artifacts, the low resolution of monitors used, the impossibility of analyzing changes of the signal obtained on the fly, etc.) can be efficiently controlled with the help of our approach. However, the goal of our study was to demonstrate that small, subtle changes in the ECG trace can be detected using the approach described. Our results prove that changes in ECG traces can be detected even in the presence of a disadvantageous signal-to-noise ratio. Whether other inaccuracies can be controlled by our approach cannot be sufficiently demonstrated based on the data available. Finally, we analyzed only one physiological baseline ECG trace as a proof-of-concept, and additionally, two pathological cases highlighting ECG morphologies. Our results do not answer whether the ability to detect subtle changes in ECG traces is preserved when different baseline ECG morphologies are used.

5. Conclusions

In this proof-of-concept study, machine learning algorithms, first of all RFs and GBMs, demonstrated high discriminatory accuracies in detecting even subtle changes in synthetically generated ECG beats despite variable degrees of artifacts present. The Boruta algorithm indicated that the features with the highest influence on adequate difference detection were the components of the ECG trace where the amplitude was artificially increased. No other parts of the ECG signal played a relevant role. Our results prove that changes in ECG traces can be detected even in the presence of a disadvantageous signal-to-noise ratio. Whether other inaccuracies can be controlled by our approach cannot be sufficiently demonstrated based on the data available. Our findings may have implications for enhancing clinical diagnosis and monitoring. Further studies are needed to validate our findings.