The blood pressure is an important factor in the diagnosis and evaluation of several diseases, such as acute myocardial infarction and stroke. This way, continuous monitorization of this parameter is crucial to a correct health evaluation. The current methods, like the oscillometric method, have some major drawbacks, that can influence the output values or even make the measurements impossible. One example is the high frequency evaluation of the blood pressure, in the standard used methods the process of measuring can take up to 3 minutes, and a waiting time is necessary between consecutive measurements.
This dissertation presents two different cuffless solution to solve those problems. One based on physical models of the human body, and the other using machine learning techniques.
In the first solution seven models that correlate pulse transit time and blood pressure, deducted by different authors, were tested to evaluate which one performed better. The tests were performed in a custom dataset acquired at Fraunhofer AICOS and in clinical environment, with two different devices (low cost device and medical grade device). The results indicate that pulse transit time can be used to track blood pressure, the developed device/method was evaluated as grade A based in the Standard IEEE 1708-2014.
The second solution it’s a proof of concept using a public database and three different machine learning methods (Random Forest, Neural Network and AdaBoost). Two sets of features are calculated from the ECG and PPG signals, one using TSFEL (spectral, frequency and time domain features) and a total of 15 custom features. The proposed method outperforms the methods presented in bibliography with mean absolute error of 3.6 mmHg and 2.0 mmHg to systolic and diastolic blood pressure respectively.
Author: Paulo Martins
Type: MSc thesis
Partner: Faculdade de Ciências e Tecnologia da Universidade Nova de Lisboa