Golden Standard or Obsolete Method? Review of ECG Applications in Clinical and Experimental Context

FIGURE 1. Original ECG records. (A) Isolated heart electrograms of various species. (B) ECG in an anesthetized Sprague Dawley male rat recorded by needle electrodes. (C) ECG in a freely moving Wistar male rat recorded by telemetry. Note the differences in the heart rate among species (A) and between anesthetized (B) and conscious (C) rats.

FIGURE 2. Milestones of ECG processing and analysis in time perspective.

FIGURE 3. Schematic representation of computer-aided ECG interpretation using feature-based technique (top) and deep learning approaches, which do not require the calculation of ECG features (bottom). Note: pre-processing steps are optional when using deep learning methods and depend on the application.

FIGURE 4. Illustration of two binary classification models: (A) straightforward artificial neural network—machine learning model providing ECG classification based on the previously derived features; (B) 1D convolutional neural network (1D CNN)—deep learning model providing automatic extraction of the features from raw ECG by convolutional layer(s) and further assignment of the input into the class by fully connected perceptron-like layers. In a 1D CNN model, the convolutional layer consists of filters, which derive important features from the input ECG. The convolution output—feature map—is usually transformed using a linear or non-linear function (in order to simplify the training process and avoid the problem of vanishing gradient) and downsampled by calculating the average (average pooling) or by selecting maximal (max-pooling) values from the feature map. The pooling procedure leads to the reduced number of model parameters and, thus, decreased computation demand.