Estimativa dos pontos de sístole e diástole para identificação de hipertensão a partir de sinais de fotopletismografia

Abstract

Cardiovascular diseases have been among the main causes of mortality and comorbidity in recent years. The main risk factor for the development of cardiovascular diseases is hypertension, characterized by the elevation of Arterial Blood Pressure (ABP), defined by systolic and diastolic heart movements. The diagnosis can be obtained through continuous monitoring by invasive and non-invasive methods. The most accurate way (gold standard) to obtain ABP comes from the cannulation technique, an uncomfortable and invasive way, subjecting the individual to an arterial perforation. On the other hand, blood pressure can be obtained by the non-invasive technique of Photoplethysmography (PPG). This signal is obtained by the process of reflection or refraction of light, with the advantage of allowing continuous monitoring of blood pressure outside the hospital environment and being implemented in low-cost wearable devices. Although there is a correlation in the waveforms among the PPG and ABP signals, both differ by their scale values providing a non-linear relationship of these signals. However, this correlation makes it possible to estimate blood pressure from the PPG signal using algorithms that find non-linear correlations between these two signals. The proposed Neural Model of Blood Pressure (MoNePS) method aims to estimate systole and diastole from the PPG signal and evaluate its predictions according to the state of blood pressure. Therefore, with the systole and diastole estimates obtained, the problem is treated as a binary classification task considering the Normotensive and Hypertensive classes. In addition, this work also investigates the performance of the method in classifying different types of blood pressure (Normotensive, Prehypertensive, Grade 1 Hypertensive, Grade 2 Hypertensive). For this, MoNePS consists of a convolutional neural network with dilatation blocks to obtain scalability in extracting PPG features and correlating the values corresponding to blood pressure in a given time window. The evaluation of the proposed method was performed using data from the public database MIMIC-III. The experimental results show that the proposed method can achieve results comparable to more complex models with fewer parameters. With the proposed method, it was possible to obtain a mean absolute error for systole and diastole of 5.02mmHg and 3.11mmHg, respectively. Furthermore, predicted outcomes subjected to multiple-class classification returned F1- Score values of up to 94% for Normotensive class. When submitted to binary classification, the model presented respective 94% and 80% for the classes of Hypertensive and Normotensive.