Search for the beginning of the diastolic rise of the photoplethysmogram using a multilayer recurrent neural network.
Gribkov Yu. Vladislav, 1
Zatrudina Sh. Rimma,1
Isupov B. Igor, 2
1 Volgograd State University, Volgograd, Russia
2 Volgograd State Medical University, Volgograd, Russia
Abstract
Photoplethysmography is an optical method for monitoring changes in blood volume in the microvascular bed. One of the main problems of processing photoplethysmograms, along with the problem of noise and interference, is the search for the moment of the onset of the diastolic rise, especially in cases where the steepness of the decline in the catacrotic pulse cycle is barely noticeable. A new method for detecting the position of the onset of diastolic rise in real time is proposed.
The method is based on the use of a multilayer recurrent neural network based on gated rekkurent unit (GRU) cells. For training and testing, a collected database of 10,000 300-ms records of PPG segments was used. The database was compiled as a result of processing by a human expert 48 records of photopretismograms with a length of 1 to 6 minutes. The input data for the network was the first derivative of the PPG signal segment, consisting of 200 elements, which corresponds to 320 millisecond recording.
The calculation graph was built using the Tensorflow 2.0.0 and KERAS libraries. The training was carried out by the method of back propagation of the error. The error was minimized in 250 iterations of the Nadam adaptive algorithm.
Verification of the positions of singular points calculated using the proposed method was carried out by comparison with the control values determined by the expert. It was demonstrated that the sensitivity of the method was about 96%. The absolute error in determining the position of the diastolic rise does not exceed 15 milliseconds, which makes the proposed mechanism sufficiently effective for practical application. The absolute error in determining the position of the diastolic rise does not exceed 10 milliseconds. The proposed method can be used in systems for computer monitoring of vascular tone.
File with abstract
Speaker
Vladislav Yurievich Gribkov
Volgograd State University
Russia
Discussion
Ask question
