The features were employed to classify various arrhythmias, such as Bradycardia, Tachycardia, Ventricular Tachycardia (VT), and Ventricular Fibrillation (VF). The outcomes of heart arrhythmias classification, based on 41 features as an input vector, are presented in Table 1. It is noteworthy to mention that several studies [24]-[29] have utilized the PhysioNet challenge database to explore and mitigate erroneous arrhythmia alerts during uninterrupted patient surveillance. Upon examination of the True Positive rate for each class in terms of classification sensitivities, the outcomes were found to be consistent with those documented in previous literature [24]-[30].
The results indicate that the Ensemble and KNN models exhibited the most efficient classification performance, achieving accuracies of 97.4% and 97.3%, respectively, among all the machine learning techniques employed. It is noteworthy that the Ensemble results in overall classification algorithms demonstrate an enhanced prediction performance compared to any individual algorithm. The sensitivity of bradycardia was high across all machine learning models, accurately classifying instances of bradycardia. This can be attributed to the straightforward nature of certain features related to pulse interval values in bradycardia.
In SVM, KNN, and Ensemble, there exists an inverse correlation between the sensitivity and specificity for VT, VF, and Tachycardia. The sensitivity for VF and Tachycardia is low due to the dominance of the positive class, while the specificity for VT is low, indicating a low true negative value and a high false negative value. The high sensitivity of VT is attributed to the large number of VT samples in the dataset. Sensitivity and precision are interrelated as they are both utilized in the calculation of true positives. However, precision is higher as a false positive is lower than a false negative, whereas sensitivity depends on the false negative.
Table 1
Cardiac arrhythmias classification results using 41 PPG features.
Classifier | Classification Report |
Arrhythmia | Sensitivity (%) | Specificity (%) | Precision (%) | Accuracy (%) |
DT | Brady | 98.5% | 93.3% | 99.8% | 93.1% |
Tachy | 85% | 85.8% | 96.8% |
VF | 82% | 85.2% | 98.1% |
VT | 96.6% | 95.7% | 93.3% |
KNN | Brady | 98.3% | 99.3% | 100% | 97.3% |
Tachy | 94% | 96% | 100% |
VF | 90.5% | 99.7% | 98.3% |
VT | 99.4% | 95% | 96.7% |
SVM | Brady | 100% | 100% | 100% | 96.7% |
Tachy | 91% | 99.3% | 96.8% |
VF | 92.1% | 99.3% | 95.1% |
VT | 98.8% | 94.5% | 96.4% |
Ensemble | Brady | 98.3% | 100% | 100% | 97.4% |
Tachy | 93% | 100% | 100% |
VF | 92.1% | 99.6% | 96.7% |
VT | 99.7% | 94.6% | 96.7% |
When evaluating the sensitivity of each class, it was found that the classification outcomes achieved through the use of just the PPG signal were superior to those reported in previous studies [24]–[29], particularly for VT and VF, as demonstrated in Table 2. In a study conducted by the authors of reference [30], the same Physio Net Challenge 2015 database was utilized to classify cardiac arrhythmias using the PPG signal, resulting in an overall accuracy of 93%, which is lower than the accuracy achieved in this study. As previously mentioned, the exclusive use of the PPG signal has the potential to significantly enhance the detection and monitoring of cardiac arrhythmias, taking advantage of its noninvasive nature. Table 2 provides a summary of the results obtained in the literature for the classification of cardiac arrhythmias using bio-signals, in comparison to the findings of this study.
The Principal Component Analysis (PCA) technique was used to reduce the dimension of the input vector and evaluate the accuracy of different machine learning algorithms versus the number of PCA components. Figure 5 shows the accuracy of different classification techniques as a function of the number of PCA components. The results show that the accuracy didn’t deteriorate significantly when the number of features in the input vector of the classifier was reduced. This could be beneficial in reducing the number of features without significantly impacting accuracy. The DT algorithm achieved the highest accuracy of 92.5% with only seven PCA components. Conversely, the SVM and KNN algorithms achieved their highest accuracy of 96.4% and 96%, respectively, using twenty-one (for the SVM) and eight PCA components (for the KNN). The Ensemble technique achieved its highest accuracy of 95.3% with nineteen PCA components. Increasing the number of PCA components beyond the specified number did not result in any performance improvement for the classification techniques.
Table 2
The classification results of cardiac arrhythmias of the current study using the Physio Net Challenge 2015 database compared with the literature.
Reference | Signal Type | Sensitivity |
Brady | Tachy | VF | VT |
Plesinger et al. [24] | ECG, PPG, ABP | 100% | 97% | 67% | 85% |
Fallet et al. [25] | ECG, PPG, ABP | 96% | 96% | 83% | 93% |
Antink et al. [26] | ECG, PPG, ABP | 100% | 100% | 67% | 90% |
Eerikanen et al. [27] | ECG, PPG, ABP | 96% | 99% | 75% | 84% |
Kalidas et al. [28] | ECG, PPG | 100% | 100% | 100% | 84% |
Caballero et al. [29] | PPG, ABP | 95% | 97% | 89% | 49% |
Paradkar et al. [30] | PPG | 88.8% | 97.6% | 50% | 88.1% |
Current study* | PPG | 98.3% | 93% | 92.1% | 99.7% |
* Using the Ensemble technique |
It can be observed that, with the exception of the DT algorithm (both pre- and post-PCA), all classification techniques may be deemed suitable for constructing the classification model. The DT algorithm's inferior accuracy may be attributed to the potential for overfitting, resulting in a complex model that fails to generalize data effectively. Additionally, if the number of waveforms is not evenly distributed across all classes, the DT algorithm may produce a biased model.
The process of selecting features was carried out through the utilization of PCA on the complete datasets. The purpose of the PCA was to examine the level of correlation between each extracted feature and the first principal components, identify the most significant features, and eliminate the irrelevant ones. The 41 features of each signal were transformed into a new set of variables, referred to as principal components, which are uncorrelated and arranged in order of significance, with the first component retaining the highest explained variance among all the original features. The explained variance denotes the information explained by a specific principal component.
The process of selecting features was carried out iteratively, as previously described. A summary of the performance of four machine learning techniques, as the number of features increased, is presented in Fig. 6. The results indicate that the selected features have marginally enhanced the classification performance. Given the similarities between Tachycardia, VF, and VT, it was necessary to exploit the dependent features for each of them by utilizing more features. The classification outcomes and the optimal features selected are summarized in Table 3.
The DT technique achieved a maximum accuracy of 94% and maintained a steady accuracy of 93.1%. On the other hand, other machine learning algorithms demonstrated high classification accuracy, with KNN achieving 98.4%, and Ensemble and SVM achieving both 98%. Figures 6 and 7 indicate that the accuracies and sensitivities of certain features increased while others decreased, which may be attributed to the hyper-parameters of the models. The inclusion of redundant features can lead to overfitting and mislead the algorithm's modeling. Additionally, the sensitivity of the classification, particularly for Tachycardia, can be affected by the inclusion of subjects with multi-arrhythmias.
Based on the findings presented in Fig. 7, it is evident that VT and bradycardia exhibit the highest sensitivity in comparison to other arrhythmias. This observation can be attributed to the larger number of VT samples as compared to Tachycardia and VF. It is noteworthy that the sensitivity, specificity, and precision of arrhythmias remain unaffected by data reduction, as they are influenced by data balancing, which ultimately determines the overall accuracy. The Ensemble classification, utilizing the same selected features as KNN and SVM, achieved an accuracy of 97.3%. Notably, there was no significant difference in accuracy between the fifteen features and thirteen features, with the latter achieving an accuracy of 98%.
The processing time of the training process is contingent upon the intricacy of the problem at hand and the complexity of the machine learning algorithm employed. The training time for the DT, KNN, and SVM algorithms was found to be comparable. In contrast, the Ensemble technique involves the amalgamation of multiple learners, thereby necessitating a training time that is directly proportional to the number of learners. Consequently, the computational complexity of the Ensemble technique is higher, and it requires a longer training time than other machine learning algorithms. However, it is noteworthy that the number of learners is not contingent upon the number of features, and the optimizer is capable of identifying the optimal number and size that balances accuracy and speed.
The results show that the KNN, SVM, and Ensemble have similar levels of accuracy, as shown in Table 3. The classifier performs better when reducing the number of neighbors. In the case of a dataset with n data points, the KNN classifier utilizes the majority voting scheme to classify new points. On the other hand, only the nearest data points are selected in case the number of neighbors is limited. Increasing the number of data points could result in a decrease in the overall performance due to the increased probability of the inclusion of a data point from other classes.
Table 3
Classification results obtained using the selected features.
Classifier | Classification Report |
Arrhythmia | Sensitivity | Specificity | Precision | Accuracy |
KNN | Brady | 98.3% | 100% | 100% | 98.4% |
Tachy | 95% | 99.7% | 98.9% |
VF | 96.8% | 99.8% | 98.3% |
VT | 99.7% | 96.8% | 97.9% |
SVM | Brady | 98.3% | 99.8% | 98.3% | 98% |
Tachy | 91% | 99.8% | 98.9% |
VF | 96.8% | 99.8% | 98.3% |
VT | 100% | 94.5% | 96.5% |
Ensemble | Brady | 100% | 99.8% | 98.3% | 98% |
Tachy | 93% | 100% | 100% |
VF | 93.7% | 99.8% | 98.3% |
VT | 100% | 95.9% | 97.3% |