Improvement of Stroke Volume Estimation with Bioimpedance Measurement by LSTM Network Approach Based on ECG during Ergometry

Abstract

Stroke volume (SV) and cardiac output (CO) estimations with non-invasive approaches like thoracic electrical bioimpedance (TEB) measurement become state of the art in clinical practice. Despite the advantages like low costs, low risk of infection and relatively easy application, there are also disadvantages like the sensitivity to movement artifacts and, electrode displacement mistakes. The bioimpedance signal acquired with a tetrapolar measurement has a relatively weak signal strength compared with another common recorded signal, e.g., the electrocardiogram (ECG). For reconstruction and filtering of the dZ/dt signal, different approaches exist like ensemble averaging (EA), scaled fourier linear combiner (SFLC), wavelet denoising and adaptive filter. We propose an artificial neural network with long short-term memory (LSTM) layer for signal reconstruction during ergometry. The LSTM network performs well compared with other algorithms, e.g., with better amplitude (C point) reconstruction. The SV estimation with the LSTM network was at least comparable or even better than the estimation based on SFLC.

GRU Neural Network Improved Bioimpedance Based Stroke Volume Estimation during Ergometry Stress Test

“…Compared to our former work [ 18 ], we could use significantly more subjects, tried to find a first method to evaluate the movement artifacts disturbed signal parts, and improve the investigation of inter-subject phenomena. An advantage of the current GRU-based approach is that it is not necessary to have an ECG QRS-complex detection as is needed for the EA, SFLC, and SNN approach.…”

“…In contrast to long-short term memory (LSTM) units, GRUs are more efficient with regard to computing time, see Figure 2 for the GRU network cell structure. Compared to our former work with LSTM networks, this is important to reduce the processing time due to the larger dataset [ 18 ].…”

“…We can calculate all hemodynamic parameters within these phases with a changed cardiovascular status due to the exercise stress test load. In the formerly published article [ 18 ], we used these breaks to evaluate our algorithms. This time we generate an estimated signal, to be precise, an estimated SV trend for the whole execution time with the breaks used as support points.…”