Estimativa da pressão arterial através de modelos de inteligência artificial no processamento de sinais de fotopletismografia

Abstract

Blood pressure measurement, and possible changes in the normality of this, it is associated with several comorbidities such as cardiovascular and cardiorespiratory diseases among others. Data reported by the Institute for Health Metrics and Evaluation (IHME) show that cardiovascular diseases are the main cause of death worldwide, killing 17.8 million people in 2017, along with 3.91 million deaths from respiratory diseases. Blood pressure is measured in two values, Systolic Blood Pressure (SBP), referring to the contraction movement of the heart, systole, where the pressure is higher, and Diastolic Blood Pressure (DBP), referring to the relaxation movement of the heart, diastole, where the pressure is less. At the present moment, the most used methods for measuring blood pressure cannot be used continuously, such as the sphygmomanometer, which contracts, usually the brachial artery, at a higher pressure than the individual’s pressure, other methods are invasive, where a catheter is required in the patient’s artery, which can cause discomfort and possible infections to the patient. This work proposes a study of techniques for the continuous and non-invasive measurement of blood pressure, through the processing of Photoplethysmography signals in conjunction with Artificial Intelligence algorithms. The research was carried out in two different datasets, one from Queensland University and the other from Federal University of Santa Maria (UFSM). Three different methodologies were developed, called Single PPG Wave, Sliding Window and Scalogram, for each method, a set of experiments was designed with the purpose of corroborating the results presented, as well as an analysis of each test. The approaches used were in the scope of regression, which estimates a floating point value, and in the classification, which seeks to situate the patient’s condition in a class. In terms of performance, this work reached the AAMI standards where the device must have an average error of less than 5 mmHg and standard deviation of less than 8 mmHg, as well as class A in the British Hypertension Society standard. The best results show that the Sliding Window method with the Random Forest algorithm reached average error and standard deviation of 0.94 ± 2.31 mmHg for SBP measurement and 0.60 ± 1.39 mmHg for DBP, among a total of 55,493 predictions.