Cardiovascular diseases represent, in the Western world, one of the first causes of death. The origins behind these problems can be multiple, but hypertension is one of the most frequent; in fact, it is estimated that it causes about 7.5 million deaths per year. Therefore, prevention becomes of fundamental importance, through the continuous monitoring of parameters such as blood pressure. As the years and the progress of technology, devices are being sought that are less and less invasive and that can be used by the single person, at home, without the need for a health worker. Recent studies have proposed various methods for the noninvasive measurement of blood pressure, such as the use of Pulse Transit Time or Pulse Wave Velocity, obtained from the electrocardiographic signal e photoplethysmography, or the use of neural networks for pressure estimation. This work is part of a project that involves the creation of a wearable device for cuffless measurement of blood pressure; the main purpose is to go to analyze different neural networks and identify the one that provides the best performance. The the task of the neural network is to provide the ABP signal at the output when it is input a signal is placed, between ECG and PPG, or both at the same time. The networks neural analyzed are the NNOE, the LSTM and the BLSTM; for each have been made several tests and finally analyzed some parameters to evaluate their performance. It is not of it is of fundamental importance that the network is able to perfectly estimate the whole shape waveform of the ABP signal, but it is more important that it be able to reproduce the peaks and valleys useful for estimating blood pressure values. The database used for training and testing networks, was built starting from the MIMIC Database, after a careful cleaning and filtering phase. The final one is composed of 46 different subjects, 27 males and 19 females, and contains numerous recordings of ECG, PPG and ABP signals, all synchronous over time. The best results were obtained with the LSTM and BLSTM networks, albeit, for how long concerns the validation with the Leave One Out method, no network provides values ​​of RMSE acceptable. This could be due to a non-optimal choice of hyperparameters, which can be modified and re-tested in future works.

AUTHOR: Anna Panerati

ADVISORS: Eros Gian Alessandro Pasero, Annunziata Paviglianiti, Vincenzo Randazzo

DEGREE COURSE: Master’s Degree in Biomedical Engineering

ACADEMIC YEAR: 2020/2021