Eficácia e acurácia de emuladores em modelos eletrofisiológicos cardíacos simulados

Abstract

Heart diseases are common nowadays, and the need for studies on electrophysiological activity has been met through computational simulations that assist in medical diagnosis. The use of statistical tools such as Uncertainty Quantification (UQ) and Sensitivity Analysis (SA) guides the development of computational models by identifying the most influential parameters. However, these tools require a large amount of data to produce reliable results. Due to the high computational cost of simulations, emulators emerge as an alternative to generate sufficient data and enable the application of UQ and SA. The objective of this work was to analyze three emulators, Polynomial Chaos Expansion (PCE), Gaussian Processes (GP), and Neural Networks (NN), regarding their efficiency and accuracy in training and predicting values of two electrophysiological models based on the ten Tusscher (2004) model. Initially, different kernels were investigated for the GPs to select the most suitable ones and compare them with the other emulators. Subsequently, the methods were evaluated based on the mean squared error (MSE) and on training and inference times, with a normalized score obtained by averaging these metrics. In general, the results show that Neural Networks presented the best balance between inference time and accuracy, especially in high-dimensional scenarios, while PCEs stood out for their fast training and good performance in specific variables. Gaussian Processes, in turn, showed competitive performance, especially in two quantities of interest, in which they achieved the lowest mean squared errors among all methods.