ECG Signal Classification Using Deep Learning Techniques Based on the PTB-XL Dataset

Figure 1. Examples of rhythm ECG signals.

Figure 2. General overview diagram of the method.

Figure 3. Diagnostic classes used in the study.

Figure 4. Convolutional network architecture. A twelve-channel ECG signal is passed through five subsequent one-dimensional convolutional layers with the LeakyReLU activation function. The results of the computation are flattened to the format of a one-dimensional vector. The results of the calculation are processed by a fully connected layer with a softmax activation function. The output value is a one-dimensional vector describing the probability distribution of the input signal belonging to each of the defined classes.

Figure 5. SincNet-based network architecture. Each channel of the 12-channel ECG signal is processed by a dedicated SincNet block. The results of each block are concatenated, flattened to the format of a one-dimensional vector, and used as an input for two subsequent fully connected layers, with LeakyReLU and softmax activation functions, respectively. The output value is a one-dimensional vector describing the probability distribution of the input signal belonging to each of the defined classes.

Figure 6. Convolutional network with the entropy features’ block architecture. The computational graph of the network is made up of two branches. In the first branch, a twelve-channel ECG signal is passed through five subsequent one-dimensional convolutional layers with the LeakyReLU activation function. In the second branch, the input signal is used to compute the vector of entropies for every channel of the signal. The results of the computations from both branches are concatenated and flattened to the format of a one-dimensional vector. The results of the calculation are processed by a fully connected layer with softmax activation function. The output value is a one-dimensional vector describing the probability distribution of the input signal belonging to each of the defined classes.

Figure 7. Confusion matrix of results for 2 classes for the convolutional network.

Figure 8. Confusion matrix of results for 2 classes for SincNet.

Figure 9. Confusion matrix of results for 2 classes for the convolutional network with entropy features.

Figure 10. Confusion matrix of results for 5 classes for the convolutional network.

Figure 11. Confusion matrix of results for 5 classes for SincNet.

Figure 12. Confusion matrix of results for 5 classes for the convolutional network with entropy features.