Optical continuous blood pressure estimation using machine learning

This master thesis deals with the continuous estimation of blood pressure
by an artificial neural network from a plethysmographic signal of an
optical sensor. For the training of the artificial neural network, single heart
cycles are extracted from the filtered and processed measurement data
and then decomposed into complex frequency components using a fast
Fourier transformation algorithm. These components are used as input
variables for the neural network with three hidden layers. Two different
data sources are used to train the network. On the one hand a public
database (MIMIC) that provides optical signals from the fingertips of
test persons and on the other hand own measurements with an optical
PALS-2 sensor at the wrist of the test persons. An invasive blood pressure
measurement is used as a reference measurement for the MIMIC database
and the continuous, non-invasive method according to Penˇ az is used for ́
the self-acquired data.
The work has shown that it is possible to determine the blood pressure
from the photoplethysmographic signal of test persons if data with similar
characteristics are available in the training data set. This is best given
within a homogeneous group of subjects. Whether it is ultimately also
possible to determine blood pressure from a photopletysmographic signal
in elderly persons and for pathological changes through a very large
reference database cannot yet be answered.