Introduction: Heart rhythm anomalies and arrhythmias are vital indicators for diagnosing and treating cardiovascular diseases. Various advancements in remote or handheld patient monitoring devices have led to the development of novel domain-specific AI architectures that enable more accurate and faster real-time ECG diagnosis. We present a new technique for decomposition of ECG signals based on Spline Variable Projection Neural Networks (VPNN) and residual auto-encoders.
Methodology: VPNN with Hermite bases demonstrated superior performance as a feature extraction layer when coupled with fully connected neural networks. Here, we use VPNN with Spline bases in regressor mode. Hence, the output of each VPNN layer is an estimation of the input. ECG segment is passed through a set of cascaded VPNN regressors, where the input of each VPNN layer is the residual of the ECG segment and the output of its preceding VPNN. VPNN regressors are initialized with different orders of spline functions and with only a small number of coefficients to prevent localization of basis functions. In such a topology, the output of each VPNN can be interpreted as a component of the input, representing specific frequency and morphological characteristics. The network is then trained using the back-propagation algorithm by minimizing the mean squared error between the input and the aggregation of components.
Results: Single-channel ECG signals from the MIT-BIH Malignant Ventricular Arrhythmia database are first split into two-second segments, and then rhythm annotations are propagated accordingly. All segments marked as noise are then discarded. In one-class anomaly detection manner, the topology is trained on all segments except ventricular tachycardia (VT) and ventricular fibrillation (VF) ones. The average normalized percent root mean squared difference (PRD) of 3.68% and 4.21% were obtained on training and test sets, respectively. Based on PRD outliers, VT/VF segments are identified with an average accuracy of 82.6%.