Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

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

Improvement of Stroke Volume Estimation with Bioimpedance Measurement by LSTM Network Approach... Current Directions in Biomedical Engineering 2019;5(1):33-36 Mike Urban*, Timo Tigges, Michael Klum, Alexandru Pielmus and Reinhold Orglmeister 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) are commonly used in clinical practice for the estimation of estimations with non-invasive approaches like thoracic the cardiovascular status of patients. CO is the product of SV electrical bioimpedance (TEB) measurement become state of and heart rate. While it is ordinary to determine the heart rate the art in clinical practice. Despite the advantages like low of a patient, e.g., with an ECG, it is challenging to determine costs, low risk of infection and relatively easy application, the SV of a patient, especially for every heartbeat. Some there are also disadvantages like the sensitivity to movement approaches exist to calculate the SV including artifacts and, electrode displacement mistakes. The thermodilution, transoesophageal doppler, and the Fick- bioimpedance signal acquired with a tetrapolar measurement method to mention just some of them [1]. Many of these has a relatively weak signal strength compared with another approaches are invasive, cost-intensive or at least not capable common recorded signal, e.g., the electrocardiogram (ECG). to estimate SV beat-to-beat. Transthoracic electrical For reconstruction and filtering of the dZ/dt signal, different bioimpedance (TEB) overcomes these disadvantages at the approaches exist like ensemble averaging (EA), scaled cost of some approach-own limitations. The main challenges fourier linear combiner (SFLC), wavelet denoising and are movement artifacts, which are the topic of this paper, the adaptive filter. We propose an artificial neural network with maximum injected current passing the descending aorta and long short-term memory (LSTM) layer for signal reliable and robust electrode placement. These topics were reconstruction during ergometry. The LSTM network investigated in previous works [2] and [3]. performs well compared with other algorithms, e.g., with TEB measurement, also known as impedance cardiography better amplitude (C point) reconstruction. The SV estimation (ICG) becomes state of the art in clinical practice. Generally, with the LSTM network was at least comparable or even an amplitude constant current with a fixed frequency in a better than the estimation based on SFLC. range of 20 kHz to 100 kHz is applied to the subject with a tetrapolar electrode configuration. The disposable outer Keywords: Bioimpedance, Neural Network, SFLC, electrodes are used for current injection, while the inner Ensemble Averaging, Stroke Volume, Cardiac Output, electrodes are used for measuring the voltage drop on the LSTM network surface. The bioimpedance has a constant part, due to the tissue resistance, and two alternating parts. One of these https://doi.org/10.1515/cdbme-2019-0009 alternating parts is the result of respiration, whereas the other part is a result of the cardiac activity (simplified model). Additionally, the cardiac related part is a result of the 1 Introduction windkessel-effect as well as the orientation of the erythrocytes and blood acceleration during a heartbeat. Stroke volume (SV) and cardiac output (CO) determination Different approaches exist to filter or estimate the bioimpedance signal and to calculate the SV. Adaptive filters are used [4] to cancel noise. They work well with noisy ______ signals, but not very well with artifacts. Furthermore, *Corresponding author: Mike Urban: Development Engineer of Osypka Medical GmbH, Germany, Berlin, and also member of ensemble averaging (EA) is used. This approach averages a Electronics and Medical Signal Processing Department, Faculty of predefined number of beats. Noise and artifacts are reduced Electrical Engineering, Technical University Berlin, 10587 Berlin, but also blurred. The dynamic depends on the number of used Germany (e-mail: m.urban@campus.tu-berlin.de). samples. More samples reduce the disturbances but lower the Timo Tigges, Michael Klum, Alexandru Pielmus and Reinhold dynamic [5]. The wavelet transformation is also used for Orglmeister: Chair of Electronics and Medical Signal Processing Department, Faculty of Electrical Engineering, Technical denoising and baseline wander removal [6]. Other University Berlin, 10587 Berlin, Germany. Open Access. © 2019 Mike Urban et al., published by De Gruyter. This work is licensed under the Creative Commons Attribution 4.0 License. M. Urban et. al., Improvement of Stroke Volume Estimation with Bioimpedance Measurement by LSTM Network Approach Based on ECG during Ergometry — 34 approaches are sensor fusion, where, e.g., 2 signals e arThis results in movement ar ifat cts. Despite this, the ECG can merged and the scaled fourier linear combiner (SFLC) [7]. be recorded, because the signal is relatively strong, compared to other signals (10-100 times larger in amplitude than the bioimpedance signal). It is very challenging to record the bioimpedance signal during such exercise. If it is possible, a 2 Methods new cardiovascular parameter could be established, for example, the slop of the SV during the exercise. The SFLC For our own setup, we used a current with 2 mA RMS uses the fundamental frequency of the RR interval and its amplitude at 50 kHz. The electrodes were placed from harmonics to reconstruct the original signal. Depending on superior to inferior alphanumerical ordered as follows. The the adaptation algorithm and coefficient, this approach results electrode A was placed 5 cm above the onset of the neck, the in a very stable signal. These coefficients also determine the electrode B was placed on the onset of the neck, the electrode dynamic behavior. C was placed on the level of the xiphoid and the electrode D was placed 10 cm below electrode C. All electrodes were placed on the left side of the patient. Electrode B and C was 2.2 Neural Network also placed on the right side (see Figure 1). The voltage drops from electrode B to electrode C and from B to C were 2 1 1 2 An approach related to the signal processing of the human measured and two bioimpedance signals were calculated. brain is an artificial neural network. These networks mimic Generally, only the electrodes on the left side are placed. For their biological counterpart. In the basic form, data is performing and sensor fusion, the bioimpedance signal is processed through the network and a learning algorithm measured twice across the thorax. If one shortens the B and C (backpropagation) adapts the inner weights of the electrodes, the disturbances will be also reduced as with a connections to produce a specific output, e.g., for supervised fusion, but also the ECG signal amplitude. The ECG is learning. This approach is part of machine learning. LSTM measured simultaneously to the bioimpedance signal (from and deep neural networks are a more specialized form. We B to C ). 2 1 use shallow neural networks separately and combine the estimated output of 25 shallow neural networks (multilayer perceptron) for an estimation of the bioimpedance signal. Additionally, we use an LSTM network to predict the bioimpedance signal, based on the ECG signal. All neural networks we used had to predict the current bioimpedance signal based on the last 200 ECG signal samples. 2.3 Procedure We performed an ergometry measurement with a healthy subject. First, a baseline was recorded for 60 s. Then the subject started to cycled 20 s, stopped for 5s and cycled again for 15 s, followed by a 5 s break. The 15/5-time interval was performed for 3 minutes. Finally, the subject had to rest for 3 minutes. We recorded an ECG from electrode B to electrode Figure 1: Electrode Placement. C and two bioimpedance signals from electrode B to 1 2 electrode C and from electrode B to electrode C . The 1 1 2 electrodes A and D were used to inject a 50 kHz current for 2.1 Ergometry the bioimpedance measurement. The breaks were included to have a reference during the exercise for the different Ergometry is used for evaluation of the cardiovascular algorithm for the waveform shape as well as for the system status and can indicate arteriosclerosis. During the calculated SV. Afterward, we processed the bioimpedance cardiac stress test on an ergometer, an ECG is recorded. signals as follows. We performed a fusion (scaled the Depending on the morphology of the ECG, especially the S- auxiliary signal and averaged the original with it) of the two T path, clinicians can assume arteriosclerosis. The st independent measurements (1 comparison signal (CS)) and measurement is done on an ergometer for heavy exercise. M. Urban et al., Improvement of Stroke Volume Estimation with Bioimpedance Measurement by LSTM Network Approach Based on ECG during Erg ometry — 35 calculated from this signal the ensemble averaged version (opening of the aortic valve), C (highest amplitude in the nd (2 CS, used 2 beats before and 2 beats after the current derivative of the bioimpedance signal, dZ/dt ) and X max signal for reconstruction). An SFLC with least mean squares (closure of the aortic valve) points in the bioimpedance rd algorithm was applied to cancel the disturbances (3 CS). signal, which are important for the determination of SV Finally, we created 25 shallow neural networks and one (specific points and SV estimation can be seen in Figure 2 as LSTM network with MATLAB. This results in three an example). We used additional signals. One was processed by the best shallow 𝐿𝐸𝑉𝑇 th 𝑎𝑥𝑚 neural network after training (4 CS), a second was 𝑆𝑉 = 𝐶 ∗ ∙ (1) 𝑍 𝑇 0 𝑅𝑅 processed by averaging 25 shallow neural network outputs where CP is a patient-specific constant (set to 1), th th (5 CS) and the last was processed by the LSTM network (6 dZ/dt is the maximum change of the impedance signal, Z max 0 CS). As training data for the LSTM network, the original is the basal impedance, LVET is the left ventricular ejection signals without the pause intervals were used. The breaks time, T is the R-R spike interval. The SV was normalized RR were used as validation data. This was done because simple to the baseline phase. Additionally, an inspection of the cross-validation is not possible. The network itself complete processed signal was performed for each method. 3 Results A good reconstruction was achieved with the LSTM network. Comparing the complete signals the resulting SV estimation of the LSTM network seems nearly like the estimation done by the SFLC method (Figure 3). The mean square error (MSE) was calculated during the break intervals between processed and original signal from electrode B to C . This 2 1 can be done because algorithms need time to adapt and will not reproduce the original signal immediately. Moreover, the MSE for the B, C and X points were calculated (Table 1). For B and X point, the time is important, whereas for C points the amplitude has to be correct. Table 1: MSE of processed signals during short breaks, specific point detection and SV estimation Signal MSE MSE B MSE C MSE X MSE SV Signal Points Points Points (norm.) 2 2 2 2 [(/s) ] [ms ] [(/s) ] [ms ] Fusion 0.23 66.89 0.236 67.2 0.0266 Fusion + EA 0.20 45.85 0.242 21.9 0.0101 Single NN 0.66 8.26 0.476 70.7 0.0174 Multiple NN 0.31 30.51 0.466 63.0 0.0239 SFLC 0.75 4.94 0.251 34.3 0.0459 LSTM NN 0.65 9.19 0.227 78.63 0.0182 The LSTM network performs well with the lowest error in C point determination (MES error 0.227 (/s) ) during breaks and a good SV determination (error 2 %). One has to reconstructs only correlated data in the bioimpedance signal consider that the reconstruction by LSTM wasn’t the best based on the ECG. After this, we compared the signal shapes during the breaks but it processed a physiological signal also during baseline (start), breaks and recovery (end) phase. during disturbed phases (compared to fusion and EA). Furthermore, we compared reconstruction of correct B 𝑑𝑡 𝑑𝑍 M. Urban et al., Improvement of Stroke Volume Estimation with Bioimpedance Measurement by LSTM Network Approach Based on ECG during Ergo metry — 36 SFLC, sensor fusion, shallow neural networks, and ensemble 4 Discussion averaging. We assume an LSTM network can learn an ECG to dZ/dt signal (bioimpedance sig.) model with good C point It seems that the LSTM network works well for a first trial. estimation. The B and X point reconstruction needs to be The signal parts during the disturbed parts are similar to the improved. The results look promising but need to be clearly SFLC approach and similar to the original signal during the validated with a larger number of subjects and additional breaks. Some signal peaks can be seen in the SV estimation measurements. The LSTM network performs similar to the of the LSTM network as a result of false bioimpedance point SFLC but with a higher dynamic. It can be used to generate a detection. During the breaks, the SFLC SV was estimated too surrogate signal for sensor fusion, to estimate the SV by high by nearly 5% compared to the original signal, because averaging or to offer the possibility of new cardiovascular of the lower dynamic. SFLC also estimated the SV in the parameters like the slop of SV during exercise. baseline phase too high. During the breaks, the LSTM network SV estimation decreases as it does in the reference Author Statement signal. Compared to SFLC, the LSTM has the additional Research funding: The author states that no funding is advantage that it only needs the ECG without R-spike involved. Conflict of interest: R. Orglmeister, Timo Tigges, annotation. The presented results are very promising with Michael Klum, and Alexandru Pielmus declare that they have respect to the strong movement artifacts in the bioimpedance no conflict of interest. M. Urban is an engineer at Osypka signal. The LSTM network can generate an ECG-to- Medical (Berlin, Germany). Informed consent: Informed bioimpedance-signal model that has more dynamic, that the consent has been obtained from all individuals included in SFLC approach. The SV estimation during the disturbed this study. Ethical approval: The research related to human phases looks similar to the SFLC result. If the generation of use complies with all the relevant national regulations, such a model is possible, one can assume, that it should also institutional policies and was performed in accordance with be possible to estimate SV directly from the LSTM network the tenets of the Helsinki Declaration. but with a complete signal, sensor fusion can be done with other bioimpedance signals or its surrogates. An averaging of SV can also be done, but one has to consider, that the B, C, References and X point detection in the postprocessing works better with an undisturbed signal. For the validation, additional [1] Garcia, X., et al. "Estimating cardiac output. Utility in the clinical practice. Available invasive and non-invasive measurements are needed with more test subjects as well as a monitoring." Medicina Intensiva 35.9 (2011): 552-561. second run with data not used for the learning process. The [2] Urban, Mike, Orglmeister, Reinhold, "Evaluation of Electrode most suitable case for the evaluation would be a beat-to-beat Setup by MRI Based Human Phantom with FEM Based SV reference that is difficult to record. Further investigations Quasi-Static Solver for Bioimpedance Measurement", IEEE, EMBC 2019 are needed to evaluate a subject-specific model with other [3] Urban, Mike, Orglmeister, Reinhold, "Surface Potential subjects. Pre-trained models should also be considered. Simulation for Robust Electrode Placement by MRI Based Approaches like SFCL and LSTM network offer the Human Phantoms with FEM Based Quasi-Static Solver for possibility of new cardiovascular status parameters like the Bioimpedance Measurement", IEEE, EMBC 2019 [4] Batra, Padma, Rajiv Kapoor, and Rakhi Singhal. "Noise adaptation rate of SV during exercise (sSV, slope of SV). Cancellation Using Adaptive Filter for Bioimpedance Signal." The simple multilayer perceptron performed bad and may International Conference on Computational Intelligence and need data alignment and normalization for a better Information Technology. Springer, Berlin, Heidelberg, 2011. performance. [5] Muzi, Michael, et al. "Determination of cardiac output using ensemble-averaged impedance cardiograms." Journal of Applied Physiology 58.1 (1985): 200-205. [6] Choudhari, Pranali C., and Dr MS Panse. "Denoising of 5 Conclusion radial bioimpedance signals using adaptive wavelet packet transform and Kalman filter." IOSR J VLSI Signal Process 5 (2015): 1-8. As a first trial, we evaluated a new approach based on an [7] Dromer, Olivier, Olivier Alata, and Olivier Bernard. LSTM network for estimate SV for disturbed signals, e.g., as "Impedance cardiography filtering using scale fourier linear a result of movement artifacts. The estimation error of the SV combiner based on RLS algorithm." 2009 Annual International Conference of the IEEE Engineering in during reference breaks was small 1.82% (normalized). The Medicine and Biology Society. IEEE, 2009. LSTM seems to be the best approach compared to others like http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Current Directions in Biomedical Engineering de Gruyter

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

Download PDF
4 pages

Loading next page...
 
/lp/de-gruyter/improvement-of-stroke-volume-estimation-with-bioimpedance-measurement-0J1NxoRKO8
Publisher
de Gruyter
Copyright
© 2019 by Walter de Gruyter Berlin/Boston
eISSN
2364-5504
DOI
10.1515/cdbme-2019-0009
Publisher site
See Article on Publisher Site

Abstract

Current Directions in Biomedical Engineering 2019;5(1):33-36 Mike Urban*, Timo Tigges, Michael Klum, Alexandru Pielmus and Reinhold Orglmeister 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) are commonly used in clinical practice for the estimation of estimations with non-invasive approaches like thoracic the cardiovascular status of patients. CO is the product of SV electrical bioimpedance (TEB) measurement become state of and heart rate. While it is ordinary to determine the heart rate the art in clinical practice. Despite the advantages like low of a patient, e.g., with an ECG, it is challenging to determine costs, low risk of infection and relatively easy application, the SV of a patient, especially for every heartbeat. Some there are also disadvantages like the sensitivity to movement approaches exist to calculate the SV including artifacts and, electrode displacement mistakes. The thermodilution, transoesophageal doppler, and the Fick- bioimpedance signal acquired with a tetrapolar measurement method to mention just some of them [1]. Many of these has a relatively weak signal strength compared with another approaches are invasive, cost-intensive or at least not capable common recorded signal, e.g., the electrocardiogram (ECG). to estimate SV beat-to-beat. Transthoracic electrical For reconstruction and filtering of the dZ/dt signal, different bioimpedance (TEB) overcomes these disadvantages at the approaches exist like ensemble averaging (EA), scaled cost of some approach-own limitations. The main challenges fourier linear combiner (SFLC), wavelet denoising and are movement artifacts, which are the topic of this paper, the adaptive filter. We propose an artificial neural network with maximum injected current passing the descending aorta and long short-term memory (LSTM) layer for signal reliable and robust electrode placement. These topics were reconstruction during ergometry. The LSTM network investigated in previous works [2] and [3]. performs well compared with other algorithms, e.g., with TEB measurement, also known as impedance cardiography better amplitude (C point) reconstruction. The SV estimation (ICG) becomes state of the art in clinical practice. Generally, with the LSTM network was at least comparable or even an amplitude constant current with a fixed frequency in a better than the estimation based on SFLC. range of 20 kHz to 100 kHz is applied to the subject with a tetrapolar electrode configuration. The disposable outer Keywords: Bioimpedance, Neural Network, SFLC, electrodes are used for current injection, while the inner Ensemble Averaging, Stroke Volume, Cardiac Output, electrodes are used for measuring the voltage drop on the LSTM network surface. The bioimpedance has a constant part, due to the tissue resistance, and two alternating parts. One of these https://doi.org/10.1515/cdbme-2019-0009 alternating parts is the result of respiration, whereas the other part is a result of the cardiac activity (simplified model). Additionally, the cardiac related part is a result of the 1 Introduction windkessel-effect as well as the orientation of the erythrocytes and blood acceleration during a heartbeat. Stroke volume (SV) and cardiac output (CO) determination Different approaches exist to filter or estimate the bioimpedance signal and to calculate the SV. Adaptive filters are used [4] to cancel noise. They work well with noisy ______ signals, but not very well with artifacts. Furthermore, *Corresponding author: Mike Urban: Development Engineer of Osypka Medical GmbH, Germany, Berlin, and also member of ensemble averaging (EA) is used. This approach averages a Electronics and Medical Signal Processing Department, Faculty of predefined number of beats. Noise and artifacts are reduced Electrical Engineering, Technical University Berlin, 10587 Berlin, but also blurred. The dynamic depends on the number of used Germany (e-mail: m.urban@campus.tu-berlin.de). samples. More samples reduce the disturbances but lower the Timo Tigges, Michael Klum, Alexandru Pielmus and Reinhold dynamic [5]. The wavelet transformation is also used for Orglmeister: Chair of Electronics and Medical Signal Processing Department, Faculty of Electrical Engineering, Technical denoising and baseline wander removal [6]. Other University Berlin, 10587 Berlin, Germany. Open Access. © 2019 Mike Urban et al., published by De Gruyter. This work is licensed under the Creative Commons Attribution 4.0 License. M. Urban et. al., Improvement of Stroke Volume Estimation with Bioimpedance Measurement by LSTM Network Approach Based on ECG during Ergometry — 34 approaches are sensor fusion, where, e.g., 2 signals e arThis results in movement ar ifat cts. Despite this, the ECG can merged and the scaled fourier linear combiner (SFLC) [7]. be recorded, because the signal is relatively strong, compared to other signals (10-100 times larger in amplitude than the bioimpedance signal). It is very challenging to record the bioimpedance signal during such exercise. If it is possible, a 2 Methods new cardiovascular parameter could be established, for example, the slop of the SV during the exercise. The SFLC For our own setup, we used a current with 2 mA RMS uses the fundamental frequency of the RR interval and its amplitude at 50 kHz. The electrodes were placed from harmonics to reconstruct the original signal. Depending on superior to inferior alphanumerical ordered as follows. The the adaptation algorithm and coefficient, this approach results electrode A was placed 5 cm above the onset of the neck, the in a very stable signal. These coefficients also determine the electrode B was placed on the onset of the neck, the electrode dynamic behavior. C was placed on the level of the xiphoid and the electrode D was placed 10 cm below electrode C. All electrodes were placed on the left side of the patient. Electrode B and C was 2.2 Neural Network also placed on the right side (see Figure 1). The voltage drops from electrode B to electrode C and from B to C were 2 1 1 2 An approach related to the signal processing of the human measured and two bioimpedance signals were calculated. brain is an artificial neural network. These networks mimic Generally, only the electrodes on the left side are placed. For their biological counterpart. In the basic form, data is performing and sensor fusion, the bioimpedance signal is processed through the network and a learning algorithm measured twice across the thorax. If one shortens the B and C (backpropagation) adapts the inner weights of the electrodes, the disturbances will be also reduced as with a connections to produce a specific output, e.g., for supervised fusion, but also the ECG signal amplitude. The ECG is learning. This approach is part of machine learning. LSTM measured simultaneously to the bioimpedance signal (from and deep neural networks are a more specialized form. We B to C ). 2 1 use shallow neural networks separately and combine the estimated output of 25 shallow neural networks (multilayer perceptron) for an estimation of the bioimpedance signal. Additionally, we use an LSTM network to predict the bioimpedance signal, based on the ECG signal. All neural networks we used had to predict the current bioimpedance signal based on the last 200 ECG signal samples. 2.3 Procedure We performed an ergometry measurement with a healthy subject. First, a baseline was recorded for 60 s. Then the subject started to cycled 20 s, stopped for 5s and cycled again for 15 s, followed by a 5 s break. The 15/5-time interval was performed for 3 minutes. Finally, the subject had to rest for 3 minutes. We recorded an ECG from electrode B to electrode Figure 1: Electrode Placement. C and two bioimpedance signals from electrode B to 1 2 electrode C and from electrode B to electrode C . The 1 1 2 electrodes A and D were used to inject a 50 kHz current for 2.1 Ergometry the bioimpedance measurement. The breaks were included to have a reference during the exercise for the different Ergometry is used for evaluation of the cardiovascular algorithm for the waveform shape as well as for the system status and can indicate arteriosclerosis. During the calculated SV. Afterward, we processed the bioimpedance cardiac stress test on an ergometer, an ECG is recorded. signals as follows. We performed a fusion (scaled the Depending on the morphology of the ECG, especially the S- auxiliary signal and averaged the original with it) of the two T path, clinicians can assume arteriosclerosis. The st independent measurements (1 comparison signal (CS)) and measurement is done on an ergometer for heavy exercise. M. Urban et al., Improvement of Stroke Volume Estimation with Bioimpedance Measurement by LSTM Network Approach Based on ECG during Erg ometry — 35 calculated from this signal the ensemble averaged version (opening of the aortic valve), C (highest amplitude in the nd (2 CS, used 2 beats before and 2 beats after the current derivative of the bioimpedance signal, dZ/dt ) and X max signal for reconstruction). An SFLC with least mean squares (closure of the aortic valve) points in the bioimpedance rd algorithm was applied to cancel the disturbances (3 CS). signal, which are important for the determination of SV Finally, we created 25 shallow neural networks and one (specific points and SV estimation can be seen in Figure 2 as LSTM network with MATLAB. This results in three an example). We used additional signals. One was processed by the best shallow 𝐿𝐸𝑉𝑇 th 𝑎𝑥𝑚 neural network after training (4 CS), a second was 𝑆𝑉 = 𝐶 ∗ ∙ (1) 𝑍 𝑇 0 𝑅𝑅 processed by averaging 25 shallow neural network outputs where CP is a patient-specific constant (set to 1), th th (5 CS) and the last was processed by the LSTM network (6 dZ/dt is the maximum change of the impedance signal, Z max 0 CS). As training data for the LSTM network, the original is the basal impedance, LVET is the left ventricular ejection signals without the pause intervals were used. The breaks time, T is the R-R spike interval. The SV was normalized RR were used as validation data. This was done because simple to the baseline phase. Additionally, an inspection of the cross-validation is not possible. The network itself complete processed signal was performed for each method. 3 Results A good reconstruction was achieved with the LSTM network. Comparing the complete signals the resulting SV estimation of the LSTM network seems nearly like the estimation done by the SFLC method (Figure 3). The mean square error (MSE) was calculated during the break intervals between processed and original signal from electrode B to C . This 2 1 can be done because algorithms need time to adapt and will not reproduce the original signal immediately. Moreover, the MSE for the B, C and X points were calculated (Table 1). For B and X point, the time is important, whereas for C points the amplitude has to be correct. Table 1: MSE of processed signals during short breaks, specific point detection and SV estimation Signal MSE MSE B MSE C MSE X MSE SV Signal Points Points Points (norm.) 2 2 2 2 [(/s) ] [ms ] [(/s) ] [ms ] Fusion 0.23 66.89 0.236 67.2 0.0266 Fusion + EA 0.20 45.85 0.242 21.9 0.0101 Single NN 0.66 8.26 0.476 70.7 0.0174 Multiple NN 0.31 30.51 0.466 63.0 0.0239 SFLC 0.75 4.94 0.251 34.3 0.0459 LSTM NN 0.65 9.19 0.227 78.63 0.0182 The LSTM network performs well with the lowest error in C point determination (MES error 0.227 (/s) ) during breaks and a good SV determination (error 2 %). One has to reconstructs only correlated data in the bioimpedance signal consider that the reconstruction by LSTM wasn’t the best based on the ECG. After this, we compared the signal shapes during the breaks but it processed a physiological signal also during baseline (start), breaks and recovery (end) phase. during disturbed phases (compared to fusion and EA). Furthermore, we compared reconstruction of correct B 𝑑𝑡 𝑑𝑍 M. Urban et al., Improvement of Stroke Volume Estimation with Bioimpedance Measurement by LSTM Network Approach Based on ECG during Ergo metry — 36 SFLC, sensor fusion, shallow neural networks, and ensemble 4 Discussion averaging. We assume an LSTM network can learn an ECG to dZ/dt signal (bioimpedance sig.) model with good C point It seems that the LSTM network works well for a first trial. estimation. The B and X point reconstruction needs to be The signal parts during the disturbed parts are similar to the improved. The results look promising but need to be clearly SFLC approach and similar to the original signal during the validated with a larger number of subjects and additional breaks. Some signal peaks can be seen in the SV estimation measurements. The LSTM network performs similar to the of the LSTM network as a result of false bioimpedance point SFLC but with a higher dynamic. It can be used to generate a detection. During the breaks, the SFLC SV was estimated too surrogate signal for sensor fusion, to estimate the SV by high by nearly 5% compared to the original signal, because averaging or to offer the possibility of new cardiovascular of the lower dynamic. SFLC also estimated the SV in the parameters like the slop of SV during exercise. baseline phase too high. During the breaks, the LSTM network SV estimation decreases as it does in the reference Author Statement signal. Compared to SFLC, the LSTM has the additional Research funding: The author states that no funding is advantage that it only needs the ECG without R-spike involved. Conflict of interest: R. Orglmeister, Timo Tigges, annotation. The presented results are very promising with Michael Klum, and Alexandru Pielmus declare that they have respect to the strong movement artifacts in the bioimpedance no conflict of interest. M. Urban is an engineer at Osypka signal. The LSTM network can generate an ECG-to- Medical (Berlin, Germany). Informed consent: Informed bioimpedance-signal model that has more dynamic, that the consent has been obtained from all individuals included in SFLC approach. The SV estimation during the disturbed this study. Ethical approval: The research related to human phases looks similar to the SFLC result. If the generation of use complies with all the relevant national regulations, such a model is possible, one can assume, that it should also institutional policies and was performed in accordance with be possible to estimate SV directly from the LSTM network the tenets of the Helsinki Declaration. but with a complete signal, sensor fusion can be done with other bioimpedance signals or its surrogates. An averaging of SV can also be done, but one has to consider, that the B, C, References and X point detection in the postprocessing works better with an undisturbed signal. For the validation, additional [1] Garcia, X., et al. "Estimating cardiac output. Utility in the clinical practice. Available invasive and non-invasive measurements are needed with more test subjects as well as a monitoring." Medicina Intensiva 35.9 (2011): 552-561. second run with data not used for the learning process. The [2] Urban, Mike, Orglmeister, Reinhold, "Evaluation of Electrode most suitable case for the evaluation would be a beat-to-beat Setup by MRI Based Human Phantom with FEM Based SV reference that is difficult to record. Further investigations Quasi-Static Solver for Bioimpedance Measurement", IEEE, EMBC 2019 are needed to evaluate a subject-specific model with other [3] Urban, Mike, Orglmeister, Reinhold, "Surface Potential subjects. Pre-trained models should also be considered. Simulation for Robust Electrode Placement by MRI Based Approaches like SFCL and LSTM network offer the Human Phantoms with FEM Based Quasi-Static Solver for possibility of new cardiovascular status parameters like the Bioimpedance Measurement", IEEE, EMBC 2019 [4] Batra, Padma, Rajiv Kapoor, and Rakhi Singhal. "Noise adaptation rate of SV during exercise (sSV, slope of SV). Cancellation Using Adaptive Filter for Bioimpedance Signal." The simple multilayer perceptron performed bad and may International Conference on Computational Intelligence and need data alignment and normalization for a better Information Technology. Springer, Berlin, Heidelberg, 2011. performance. [5] Muzi, Michael, et al. "Determination of cardiac output using ensemble-averaged impedance cardiograms." Journal of Applied Physiology 58.1 (1985): 200-205. [6] Choudhari, Pranali C., and Dr MS Panse. "Denoising of 5 Conclusion radial bioimpedance signals using adaptive wavelet packet transform and Kalman filter." IOSR J VLSI Signal Process 5 (2015): 1-8. As a first trial, we evaluated a new approach based on an [7] Dromer, Olivier, Olivier Alata, and Olivier Bernard. LSTM network for estimate SV for disturbed signals, e.g., as "Impedance cardiography filtering using scale fourier linear a result of movement artifacts. The estimation error of the SV combiner based on RLS algorithm." 2009 Annual International Conference of the IEEE Engineering in during reference breaks was small 1.82% (normalized). The Medicine and Biology Society. IEEE, 2009. LSTM seems to be the best approach compared to others like

Journal

Current Directions in Biomedical Engineering – de Gruyter

Published: Sep 1, 2019

Keywords: Bioimpedance; Neural Network; SFLC; Ensemble Averaging; Stroke Volume; Cardiac Output; LSTM network

There are no references for this article.