Unsupervised Learning to Analysis of Population of Models in Computational Electrophysiological Studies

Unsupervised learning is a helpful tool for processing experimental data. These types of machine learning algorithms can address problems of dimensionality reduction, clusterization, anomaly detection, visualization, and many others. We propose that unsupervised learning may also be useful for the analysis of the results of computational experiments. This proof-of-concept paper is about the application of several unsupervised learning algorithms for computational studies of the electrical activity in human atrial cardiomyocytes. A dataset for analysis was generated with a population modeling approach, which is widely used in cardiac modeling. At first, the signals generated by a model population were processed with the manifold learning and clusterization methods. Then, the dataset was processed with the deep autoencoder neural networks, and the latent space of the autoencoder was used as a tool for comparison of the two ionic models of the human atrial cardiomyocyte. Applied methods allowed us to classify the model solutions according to their physiological meaning, to visualize the model behavior, and to find special types of model solutions which would be almost impossible to notice via manual analysis.