ECG synthesis from separate wearable bipolar electrodes D. Farotto, L. Atallah, P. van der Heijden, L. Grieten
Abstract— Compared to cabled ECG devices, the use of wearable patches to reconstruct ECG offers a more comfortable alternative for continuous monitoring, especially for patients at home. In this work, we investigate the feasibility of synthesizing a 3-lead ECG signal from 3 separate wearable and wireless patches. We also investigate the effect of their orientation on the synthesized signal. We conduct an experiment on healthy subjects and show the ability of this method to provide a similar ECG signal to the reference, in terms of matching the overall signal pattern. However, a more detailed study would be needed in order to investigate the ability of this method to identify critical conditions for vulnerable subjects.
I. INTRODUCTION The electrocardiogram (ECG) is a widely accepted technique to monitor cardiac activity. Research in the last few decades has led to significant improvements in the acquisition and analysis of ECG. It has also caused the reduction of cable clutter /device obtrusion leading to improved patient comfort. An important development was the use of techniques to obtain ECG (normally a 12-lead) from a reduced lead set , . This allowed the reduction of the application time as well as lowering the price of such devices. For long term monitoring, however, a more practical solution consists of using 1-lead ECG monitoring with wireless patch . A wearable ECG patch contains wireless Short Distance Bipolar Electrodes (SDBE) which measure the ECG signal from two closely placed skin electrodes. Previous studies investigated the optimal location and intra electrode distance for such a patch ,. These studies showed that a 5cm distance between electrodes provided acceptable signal-to-noise ratio. Previous studies also investigated the possibility of synthetizing the standard ECG from a limited lead set. The most common method used was based on linear regression. However, deterministic approaches and neural networks were also used ,. When using SDBEs, the method used mostly was linear combination based on body potential maps ,,. Studies  and  showed high cross-correlation and low root main square error between the original standard ECG and the synthetized ECG. However, these studies did
not consider the synthesis of ECG from separate patches, which introduces many challenges that are not covered by acquiring ECG from body potential maps. These include the fact that the separate patches are not normally connected to a common reference, and can have issues with synchronization. In addition to that, patches normally have a battery which poses limits on both recording time and sampling rates used. The aim of this study is to investigate the synthesis of a 3lead ECG from 3 separate wireless patches using a linear reconstruction method. Positive results would confirm the applicability of the methodology to real home-monitoring situations. The method would help in providing a patientfriendly method for continuous monitoring of 3-lead ECG, especially for mobile patients who want to retain their freedom of movement. II. METHOD A. Subjects and apparatus Six male healthy volunteers with an age between 24 and 30 were included in this study. Informed consent was obtained prior to the experiment. . The following devices were used for measurement: -3 ECG necklaces, provided from by Holst, used as wearable and wireless SDBEs, with a sampling frequency of 256 Hz. An example of the device is shown in Figure 1. -Standard 12 lead ECG device used with a patient monitor, obtained with a vacuum electrode system and with a sampling frequency of 100Hz. The devices were used concurrently. The patches were located as in Fig. 2 with an intra-electrode distance of 5.5 cm.
Resrach supported by Philips Research. D. Farotto is with Philips Research, Eindhoven, Netherlands; e-mail: ([email protected]
). L. Atallah is with Philips Research, Eindhoven, Netherlands; e-mail: ([email protected]
). P. van der Hijden is with Holst, IMEC, Eindhoven, Netherlands; e-mail: ([email protected]
). L. Grieten is with Holst, IMEC, Eindhoven, Netherlands; e-mail: ([email protected]
978-1-4244-9270-1/15/$31.00 ©2015 IEEE
Figure 1: The Holst necklace (http://www.holstcentre.com)
Lead I [V]
Lead I [V]
Figure 2. Schematic location of the electrodes during the experiment. The red ovals indicate location and orientation of the patches, the blue dots the respective position of the electrodes for the standard ECG.
Lead II [V]
Lead II [V]
1000 0 -1000 2000
Lead III [V]
Lead III [V]
0 -500 2000
1000 0 -1000 50
1000 0 -1000 50 55
50.4 50.6 time [s]
Figure 3. Results: The left side shows an example of the measured ECG (blue) from the reference and of the syntethized ECG (red). The right side shows the signals in more detail.
The experiment was performed in two different configurations, with the individual patches placed in a vertical then in a horizontal direction as in Fig. 2. This was done to determine if the orientation of the patches would affect the synthesized ECG. The subjects were asked to sit still for 4 minutes for each configuration. The subjects were later asked to move their upper body (in slow motions) while wearing the patches and having the reference electrodes to observe the effect of motion on both modalities.
filter, bandwidth 0.1-40 Hz). Signals from the patches were then down-sampled to 100 Hz in order to have all data at the same sampling frequency.. The signals were then manually synchronized and segmented in two parts: the first 10% was used to derive a linear relationship between the standard ECG and the data from the patches; the rest 90% was used for quality evaluation. Using linear regression, the standard ECG constitutes the target for the synthesis, as following:
B. ECG synthesis For data processing purposes, MATLAB (MathWorks Inc.) was used. Each signal was filtered (3rd order Butterworth
ECG(t)i (t) ∙i,k
Where ECGk indicates the k lead from the standard ECG;
indicates the error; i indicates the signal from patch i, indicates the coefficient for patch i and it is estimated using the least-squares method for every lead k. C. Evaluation of performance To assess how well we reconstructed the ECG, we used two methods: the root main square error (RMSE) and cross correlation between the synthesized signal and the reference for the 3 ECG leads. III. RESULTS The results of 3-lead synthesis from three SDBE for one subject are shown in Fig. 3. We can see that in general, the temporal detail is captured, and the synthesized signal from the 3 patches is similar to the standard ECG signal. However, we see a lower QRS peak amplitude in the signal synthesized from the wearable patches. In addition to that, the p-wave is not clearly evident in the synthesized signal in most PQRST complexes for all the subjects. Fig.4 shows the performance indices for the two configurations, averaged for all subjects during the quiet periods without motion. Overall, the RMSE is similar for horizontal and vertical patches for all 3 leads. The cross
correlation values only show differences for lead 1, and are quite similar for leads 2 and 3. During motion, we observed both the reference and the signals synthesized from the patches. The signals were quite noisy and hardly showed the P and T waves, which made it difficult to compare the synthesized ECG to the reference signal. This why we limited our results (Fig 4.) to the quiet periods without motion. IV. DISCUSSION The main objective of this study was to investigate the feasibility of the approach aiming to synthetize 3-lead ECG from a combination of 3 separate wireless patches. In order to accomplish this, an experiment was conducted and signals from three patches were linearly combined and compared with the standard ECG. Fig. 3 shows that the synthetized ECG can identify the main shapes of the ECG signal. However, we noted an amplitude difference between the 2 signals although temporal features were well matched. Regarding the effect of the orientation of the wireless patches, it did not have a major effect on the synthesis, as long as a training period was used to derive the linear transformation between the 3 patches and the reference. From a synthesis point of view, the method we investigated used linear reconstruction, optimizing an overall mean
500 400 300 200 100
0 Lead I
80 60 40 20
Figure 4. Results: RMSE and cross-correlation for vertical (left side) and horizontal patch configurations (right side). This was calculated for the periods without motion.
squared error. This method provides an overall match but might not pick up signal details accurately, which is probably what led to the p-wave not being well-captured. It is difficult to compare the results of this study directly to previous studies who used connected ECG to a common ground and did not use separate bi-polar patches. The procedure used in study  for example consisted in using personalized patches positions which allowed an optimal synthesis per subject (for a 12 lead ECG). The method allowed better results in terms of RMSE error but the setup was quite different. This work presented a first step in understanding the performance of separate patches for synthesizing a 3 lead ECG. More experimental work is required to study ECG synthesis across a larger number of subjects, with a higher sampling frequency than the one used, aiming to better capture signal details. The effect of motion also needs to be analyzed further. The ability of this method to identify ECGsignal abnormalities, such as ST elevations and arrhythmias would also be very important to investigate if this method was to be used for the home monitoring of patients. ACKNOWLEDGMENT We would like to thank all participants in the experiment as well as the team from Ziekenhuis Oost-Limburg, in Genk, Belgium for their help with data collection. REFERENCES  Atoui, H., J. Fayn, and P. Rubel. "A neural network approach for patient-specific 12-lead ECG synthesis in patient monitoring environments." Computers in Cardiology, 2004. 2004. 161-164.
 Baig, M. M., H. Gholamhosseini, and M. J. Connolly. "A comprehensive survey of wearable and wireless ECG monitoring systems for older adults." Medical & Biological Engineering & Computing (Springer Heidelberg) 51, no. 5 (#may# 2013): 485495.  Cao, H. S., H. M. Li, L. Stocco, and V. C. M. Leung. "Wireless Three-Pad ECG System: Challenges, Design, and Evaluations." Journal of Communications and Networks (Korean Inst Communications Sciences (k I C S)) 13 (2011): 113-124.  Dower, Gordon E., Andrew Yakush, Sami B. Nazzal, Roy V. Jutzy, and Cynthia E. Ruiz. "Deriving the 12-lead electrocardiogram from four (EASI) electrodes." Journal of Electrocardiology 21, Supplement (1988): S182 - S187.  Horacek, B.Milan, James W. Warren, Dirk Q. Feild, and Charles L. Feldman. "Statistical and deterministic approaches to designing transformations of electrocardiographic leads." Journal of Electrocardiology 35 (2002): 41-52.  Puurtinen, M., J. Hyttinen, J. Viik, P. Kauppinen, and J. Malmivuo. "Estimation of ECG signal of closely separated bipolar electrodes using thorax models." Proceedings of the 26th Annual International Conference of the Ieee Engineering In Medicine and Biology Society, Vols 1-7 (Ieee) 26 (2004): 801-804.  Puurtinen, Merja, Jari Viik, and Jari Hyttinen. "Best Electrode Locations for a Small Bipolar ECG Device: Signal Strength Analysis of Clinical Data." Annals of Biomedical Engineering (Springer US) 37 (2009): 331-336.  Tomasic, I., S. Frljak, and R. Trobec. "Estimating the Universal Positions of Wireless Body Electrodes for Measuring Cardiac Electrical Activity." IEEE Transactions On Biomedical Engineering (Ieee-inst Electrical Electronics Engineers Inc) 60 (2013): 3368-3374.  Trobec, R., and I. Tomasic. "Synthesis of the 12-Lead Electrocardiogram From Differential Leads." Information Technology in Biomedicine, IEEE Transactions on 15 (2011): 615-621.