Estimação de parâmetros do modelo do circuito equivalente do sistema cardiovascular humano usando deep learning

Abstract

For patients with severe heart diseases, heart transplantation is still the best treatment option. However, the so-called Ventricular Assist Devices (VADs) have been used successfully to support the pumping of the cardiac muscle to meet the needs of the human cardiovascular system (CVS). The so-called lumped parameter (0D) models are of great importance for computational simulations of hemodynamic variables (HVs), either using CVS models or VAD models, making it possible to analyze the performance of different operation modes even before implanting the device in the patient. Furthermore, specific models for a given patient allow the tuning of control systems to be carried out according to the clinical situation of that patient. It is known that the parametric estimation process of such models requires patient data and the HVs of interest are not always. Thus, using HVs preferably obtained through non-invasive techniques and those considered common in the medical-hospital environment should be sought. This work investigates the feasibility of implementing a parametric estimation process of a 0D model of the human CVS for specific patients. For this purpose, deep learning techniques are used, having only the arterial systemic blood pressure signal as input signal since it is can be obtained using non-invasive methods. The synthetic database was generated and divided in 3 subsets: training, validation and test. This subsets was used for training the neural network models.The sensitivity function is calculated to investigate the influence of the CVS parameter variation on all HVs since this correlation is directly related to the accuracy of the estimated parameter values. The results highlight the very low sensitivity of systemic pressure for certain parameters. This fact impairs their estimation and confirms the need to add more HVs as inputs to the estimator. The sensitivity study also highlights that some parameters’ variation does not significantly influence any of the HVs, impairing the entire estimation process for a specific patient.