Analysis of epicardial electrograms using artificial intelligence

One of the most effective electrophysiological methods for studying the myocardium today is the method of using microelectrode arrays, which is characterized by high spatial resolution of recording extracellular potentials. Difficulties in analyzing bioelectric potentials recorded from living objects of study (cell, tissue, organ) by direct methods lie in the instability of the shape, amplitude and frequency of the recorded bioelectric potentials depending on the experimental conditions, as well as in the presence of electrical noise and artifacts. There is a need to constantly monitor the parameters of the signal processing algorithm over multiple short time intervals, followed by careful verification of the result. Taking into account the high recording sampling rates of modern measuring equipment and the impressive volumes of output data, the lifted to use artificial intelligence algorithms to solve these analytical problems becomes obvious. In addition, the use of artificial intelligence methods has great prospects for identifying predictors of life-threatening arrhythmias in cardiac electrograms during experimental modeling of these conditions. The electrograms involved in the study were obtained by multielectrode mapping with flexible arrays, including 64 recording electrodes, from the epicardial surface of isolated perfused rat hearts. The moments of activation on the electrogram graph mean the points of maximum steepness of the potential decline, which correspond to the moments of the appearance of action potentials on the membranes of cardiomyocytes, that is, tissue excitation. Analysis of the frequency of occurrence of activation moments on one electrode or the sequence of occurrence on several electrodes within the arrays allows one to evaluate such parameters of the heart as its pacemaker activity and electrical conductivity of the myocardium. As part of the study of the bioelectrical activity of the heart, a promising direction is the use of artificial intelligence methods to automate the analysis of electrograms recorded from the surface of the epicardium. The presented work describes the creation of a software package for analyzing electrograms of isolated hearts of small rodents, the main part of which is a segmenting neural network for localizing moments of myocardial activation based on UNet architecture. The choice of this architecture is due to its effectiveness in image segmentation tasks, which is especially important for identifying structures in cardiac electrograms. UNet architecture is characterized by the presence of convolutional layers for feature extraction and a decoder for accurate reconstruction of spatial information. This makes it an excellent choice for medical data segmentation tasks, such as electrograms, where prescision and recall are critical. However, as mentioned earlier: UNet from the original paper is intended for image segmentation, and therefore the neural network was adapted for the analysis of one-dimensional signals. Due to the small amount of labeled data, crossvalidation was carried out to measure the quality of the model; it was evaluated on eight folds. The success of segmentation is measured by Fl metric, which is the harmonic mean between precision and recall. In this context, an Fl value of around 0.77 indicates the model’s ability to accurately identify and localize moments of activation in a heart. The goal of the work is to create software that includes the following functionality: creating a data set for training, validation and testing, training a model; creating and editing markup. Taken together, this will allow automatic localization of activation moments in epicardial electrograms. Thus, the software package we have developed ensures the identification and precise determination of the desired moments of activation, which facilitates further analysis of bioelectrical activity and increases the efficiency of research in the field of cardiology, including due to the possibility of processing big data. In general, the developed software package represents a promising solution for automating the analysis of epicardial electrograms using a segmenting neural network based on UNet architecture and related algorithms.