Samples of the textile materials used in this study were kindly provided by the Department of Electrical and Electronic Engineering, University of Cagliari, Cagliari, Italy. Fabrication process is described by Panni et al. [11]. Briefly, textile electrodes were made by treating conventional fabrics with a conductive solution of PEDOT:PSS dispersion Clevios PH 500 (Heraeus Clevios—Germany), second dopant was glycerol 33%. Woven fabrics were used, and immersed for at least 48 h at room temperature in the polymer solution. Fabrics were then taken out from the solution and drained off to remove the solution in excess. Samples were annealed, for both water and dopants to evaporate in order to avoid deterioration of the fabric mechanical properties. The conductive fabrics used in this study were: cotton, cotton–polyester (65% cotton, 35% polyester), lycra, polyester; additionally, we included a commercial silver-plated nylon fabric Shieldex®Med-Tex P130 (Statex—Germany) [23].
Figure 1 shows five electrodes manufactured following the process described by Pani et al. [11]. The fabrics were cut into pieces of 20 mm × 20 mm, which were sewn to a non-conductive synthetic leather with silver-coated yarn to obtain greater rigidity. The size of the electrodes, which is acceptable for ECG monitoring, was chosen to ensure reproducibility of the customized fabrication process. The use of a layer of rigid synthetic leather allowed to improve the contact between the electrode and the skin ensuring a uniform pressure, which is especially beneficial in the case of textile electrodes.
Finally, we fixed a metallic snap fastener to the synthetic leather and interfaced them with the same conductive yarn. In such a way, the snap fastener remained at the rear of the electrode without getting in touch with the skin. Figure 2 shows a closer view of the final aspect of the electrode, its structure, and components. The snap fastener was used to connect the electrodes to the ECG leads. In the experiments, we utilized Ag/AgCl disposable electrodes ref 2228 (3M, Germany) as the reference electrode on the ECG recording arrangement. This work focuses on electrode-skin interactions, other evaluation tests to characterize the physical properties of the electrodes were not conducted.
Pani et al. [11] reported the values of conductivity for cotton and polyester. In this work, we calculated the conductivity for cotton–polyester and lycra as the product between thickness and surface resistance, considering the fabrics as thin films with uniform surfaces, commonly reported as two-dimensional entities. Med-Tex P130 conductivity was not calculated as per its plated surface, it is intended to obtain silver ionic release for wound care, skin disorders, skin irritations, burn victims, not for uniformed conductivity.
Figure 3 shows optical micrographs of each type of fabric where is possible to appreciate the different types of weave. Optical micrographs were acquired using a 10× objective and an upright microscope (Eclipse Ci; Nikon).
Our test protocol was previously approved by the Committee of Health Research Ethics of Universidad Pontificia Bolivariana (Colombia), located therein in the document R.R. N 80 17-12-2008. Data were obtained from 8 healthy, slim build individuals between the ages of 18 and 30, four for each test (two men and two women). Our interest lied in the number of repeated measurements of each type of electrode rather than in a large number of individuals. Even though the evaluations were carried out with four participants, the noise was measured in eight individuals: four individuals chosen originally for the noise test and other four resulting from the first long-term performance measurements since both tests followed the same protocol. We set the experiments in an in-paralleled electrode configuration. We replaced the disposable electrodes at every test to avoid adding a new variable into the experiments.
The volunteers were informed about the protocol to which they would be subjected. They were asked to be at rest for a period of 30 min before the test to homogenize their body conditions. Then, the area to be measured was shaved and cleaned with alcohol to improve the adhesion of the electrodes to the skin. An elastic waistband was used to attach the textile electrodes to the skin. Neither adhesives nor electrolytic gel were used.
We used an R language based platform known as Rwizard [24] to perform all the statistical analysis.
The main electronic equipment that we used in the study was:
-
A virtual instrument, composed by a two channels USB oscilloscope and a function generator, Handyscope HS5 (Sneek, The Netherlands).
-
A device to measure low voltages, currents, and power, Cassy Lab (LD DIDACTIC GmbH, Hürth, Germany).
-
A switching circuit which is driven by a microcontroller to measure the power terminals at different points of the circuit.
-
An acquisition card based on an EVM ADS1298 device, low-power, 24-bit, simultaneously sampling, eight-channel front-end for ECG and EEG applications (Texas instrument, Texas, EEUU).
-
A laptop to set up the electronic systems and to record the data.
Measurement strategies are described in the following sections:
Contact impedance measurements
Contact impedance refers to the impedance at the skin-electrode interface. This test intends to quantify the property of the electrode-skin contact to oppose time-varying electric current produced by the material under test. We selected single and double dispersion Cole impedance models of first and second order to represent this parameter [25, 26]. We set a variable AC source at 5 V peak to peak (\( 5~V_{pp} \)) to sweep in a range of 0.1 Hz–10 kHz. Although the spectral components of ECG signals do not exceed 150 Hz, it is strategic to measure high frequencies to tune the models [27].
We used a variation of the method reported by Xie et al. [12]. The procedure involves determining the response of the electrodes when a sinusoidal voltage source is swept in frequency. This method is effective in measuring absolute magnitude of the impedance, however, does not allow the discrimination of the resistive and reactive components. To find such components individually, we performed the procedure based on the scheme in Fig. 4a. Instead of using a multimeter, we used a digital oscilloscope to estimate the magnitude and phase components of the contact impedance.
\(V_g\) supplies AC signal (\( 5~V_{pp} \)) to the circuit through the electrode 3; \(V_{e2}\) corresponds to the voltage measured at electrode 2; the voltage \(V_r\) at the reference resistance \(R_{ref}\) is calculated with \(V_g\) as the reference, and \(V_{21}\) satisfies \(V_{e2} - V_r\). \(V_r\) and \(V_{e2}\) are measured in the phasor form (magnitude and phase). The impedance is calculated as \(Z_{sum} = Z_{contact} + Z_{SB12}\), where \( Z_{contact}\) represents contact impedance (skin/electrode) of a single electrode 1 and \(Z_{SB12}\) represents the impedance of the subcutaneous tissue between the electrodes one and two. It can also be calculated by the expression:
$$\begin{aligned} Z_{sum}=Z_{contact}+Z_{SB12}=\frac{V_{21}}{I} \end{aligned}$$
(1)
where I is the current in the circuit and can be calculated as \(I=\frac{V_r}{R_{ref}}\) (both values known). The circuit of Fig. 4b. allows determining the impedance \(Z_{12}\), which satisfies \(Z_{12}=2Z_{contact} + Z_{SB12}\). \(V_g\) supplies AC signal (\( 5~V_{pp} \)) to the circuit through the electrode 2; the voltage \(V_r\) is measured at the reference resistance \(R_{ref}\) and is obtained by using \(V_{21}=V_g - V_r\). Thus, the impedance is calculated as:
$$\begin{aligned} Z_{12}=2Z_{contact}+Z_{SB12}=\frac{V_{21}}{I} \quad\text{where}\; I=\frac{V_r}{R_{ref}} \end{aligned}$$
(2)
Finally, contact impedance is calculated as \(Z_{contact} = Z_{12}-Z_{sum}\).
Lissajous figures
Given an input signal x(t) and a phase-shifted output signal y(t) such as:
$$\begin{aligned} x(t)= \; & {} X_0sin(\omega t) \nonumber \\ y(t)= \; & {} Y_0sin(\omega t+\theta ) \end{aligned}$$
(3)
the Lissajous figure is generated when plotting \(x(t)\ \text{vs}\ y(t)\) signals. The intersection of the figure with the y axis is identified and named \(y_0\). The phase shift between the waveforms is calculated by the expression:
$$\begin{aligned} \theta =arcsin \left( \frac{y_0}{Y_0}\right) \end{aligned}$$
(4)
Correlation analysis
Cross-correlation is a measure of the similarity between two series as a function of one relative to the other. The cross correlation between a discrete input signal x(n) and an output signal y(n) is given by the expression:
$$\begin{aligned} r_{x,y}(l)=\sum _{n=-\infty }^{\infty }x(n)y(n-l) \end{aligned}$$
(5)
where l represents a shift in discrete time between signals. The aim is to find the value of l in such way that the maximum correlation between signals in obtained. From this value, the phase shift in degrees can be calculated from the equation:
$$\begin{aligned} \theta =\frac{360\cdot l\cdot F}{F_s} \end{aligned}$$
(6)
where F corresponds to frequency of the original wave and \(F_s\) to the sample frequency from which the signals were acquired.
We determined voltages in the phasor form from their waveforms at the frequencies of interest. We designed a high-pass digital filter for eliminating the DC offset (Filter Designer App, Matlab, Mathworks, Inc.), with the following parameters: FIR, cut-off frequency = 0.05 Hz, order = 2000. Filter was applied off-line to both kind of signals (acquired from textile and reference electrodes), thus delays affected equally. The magnitude was calculated by obtaining the peaks of each wave. The phase was calculated using two techniques:
In order to compare the absolute impedance values of textile and disposable electrodes, we converted each data set of curves into a single scalar value. We have used the AUC score as a comparison parameter since the behavior of the magnitude of the impedance in the magnitude–frequency plots, decreases monotically with the increase of the frequency for all the samples. The use of the AUC score as a comparison criterion for spectral curves was used previously by Sarbaz et al. [28].
We used a multifactorial ANOVA (either two-way or repeated variables) in cases where the statistical assumptions of normality and homoscedasticity were applicable. For those cases when the assumptions were not satisfied, we performed nonparametric tests, such as Kruskal–Wallis and Wilcoxon. The main factors evaluated were: the material of the textile electrode (cotton, cotton–polyester, lycra, silver-plated nylon, and polyester), and their behavior relative to the reference (textile electrode vs Ag/AgCl electrodes). The measurement scheme for this test is shown in Fig. 5.
We selected three different points of each leg to perform the measurements, to take into account the local variations of the skin. The tests are performed on the legs due to the ease of locating several electrodes for simultaneous measurements. In this way the probability of presence of motion artifacts and the interference from other bioelectric signals (such as ECG and breathing signals) is reduced.
Polarization measurements
The electrode polarization is a consequence of an alteration of the charge distribution in the skin-electrode interface, and causes a baseline drift or DC potential/offset in ECG signals. Normally in practice, measuring ECG signals requires at least two electrodes connected differentially to an instrumentation amplifier that reduces the effects of common mode interference. ECG trace must be amplified and DC potential reduced. Electrochemical phenomena at the skin cause variations, as polarizations, on the skin-electrode interface, which results in interfering and modifying signals that are added to the desired ECG signal, although the characteristics of the electrodes are the same. Polarization at the skin-electrode interface was calculated by measuring a DC potential in open circuit at the terminals of a pair of electrodes attached to the skin. The potentials were registered once the patient remained one minute motionless to avoid instabilities in the skin-electrode interface, product of involuntary biomechanical movements. The skin-electrode interface is the largest source of interference due to polarization potentials. Polarization potentials are normally in the order of millivolts; however, when values exceed such order at the presence of action potential variations, the output of the amplifier is saturated, making the ECG signal difficult to extract and polarization potentials difficult to eliminate.
We set two independent acquisition channels (Cassy Lab, Label Didactics., Ltd.) to perform simultaneous measurements of DC potentials. The measurements were carried out both, in textile and reference electrodes in the same muscle group. We programmed a series of measurements by using four electrodes (two textiles electrodes and two Ag/AgCl electrodes), as reported by Rattfalt et. al. [19]. We used three different points of each leg, (sampling frequency = 10 sps), during approximately 30 min.
We calculated DC potentials for each type of electrode as the mean absolute difference between two consecutive samples. We used the standard average exchange ratio \(\bar{X}_i\) suggested by Rattfalt [18, 19].
$$\begin{aligned} &\bar{X}_i=\frac{\sum _t|X_i(t)-X_i(t+1)|}{N-1} \quad t=0,1,2,...,N-1 \nonumber \\ &\bar{X}=\frac{\sum _i\bar{X}_i}{n} \end{aligned}$$
(7)
where \(i\) denotes each particular individual; \(n\) the total number of individuals for each electrode type, and \(N\) the total number of samples. It is necessary to guarantee that the patient is motionless to avoid muscle signals product of involuntary movements.
We used the interquartile range to eliminate the outliers in each set of observations. Each measurement series became a datum representing the average behavior of the electrode through the time. A two-way ANOVA analysis was used. We selected the type of electrode as the factor, the assumptions of normality (Kolmogorov–Smirnov, p = 0.2051) and homoscedasticity were satisfied (Levene test, p = 0.1149). The measurement scheme for this test is shown in Fig. 6.
Noise measurements
These experiments intended to quantify the noise level due to external interference, biological signals different to ECG, artifacts, and measuring equipment. We performed simultaneous measurements of the textile and commercial Ag/AgCl electrodes. The experiments consisted in capturing the same 1-lead ECG using different pairs of textile electrodes. We performed the measurements using lead II, as suggested by Takamatsu [29]. The acquisition process was conducted for a period of five minutes, where textile electrodes were attached to the skin by an elastic waistband. Figure 7 depicts the location of the electrodes and the connection to the electronics acquisition card EVM ADS1298 (Texas Instruments).
We designed a digital filter (Filter Designer App, Matlab, Mathworks, Inc.) for removing undesired components from the power supply and their corresponding harmonics (2 stopband filters, FIR filters, stop frequencies = 60 and 120 Hz respectively, windowing method = Kaiser, \( \beta =0.5 \), order = 150, broadband = 10 Hz) and attenuating the frequency components out of the range of cardiac signals (passband filter, FIR filter, band pass = 0.05–150 Hz, windowing method = Kaiser, \( \beta =0.5 \), order = 150), as is suggested in [30]. We performed three methods to analyze the data: noise power, cross-correlation coefficient, and segmentation.
Noise power quantifies the magnitude of the signal eliminated in the filtering process. The aim of this procedure is to identify which type of electrode has greater affectation by external interferences, biological noise, artifacts of muscle movement and breathing. The process involves determining the difference between the original and the filtered signal to calculate the average power.
$$\begin{aligned} &E=|ECG_{original}-ECG_{filtered}| \nonumber \\ &\bar{P}=\frac{1}{N}\sum _{i=0}^{N-1}E^2 \end{aligned}$$
(8)
where E is the absolute value resulting from the difference between the original and filtered signals. \(\bar{P}\) is the noise power and N represents the number of samples.
The second method is the Pearson cross-correlation coefficient. Since the cardiac signals were recorded simultaneously with both, the textile and disposable electrodes, in the same area of the volunteer’s body, we expected two morphologically identical signals. However, they suffered a potential drift that was removed using a digital high pass filter described above. The normalized cross-correlation provides a value that expresses the similarity of two signals in terms of morphology; therefore, low values of the cross-correlation index suggest a large effect of noise on the ECG signals recorded by different electrodes.
The third process involves the use of a segmentation algorithm, which detects and quantifies complete P–Q–R–S–T waves. It uses the continuous wavelet transform, discrete wavelet transform, and Pan and Tompkins algorithm for the classification of the ECG signal, as reported by Bustamante et al. [31]. The error rate is calculated by dividing the number of complete ECG segments registered with the experimental material, against the number of ECG segments captured simultaneously with Ag/AgCl commercial electrodes.
Long-term performance
The performance of the electrodes over time is affected by the wear of the material. We evaluated the degree of deterioration of the textile electrode quantifying its capacity to record complete ECG complexes that are morphologically similar to those recorded by Ag/AgCl electrodes. The signal acquisition process was the same as described in the noise measurement section. We evaluated each type of electrode (cotton, cotton–polyester, lycra, polyester and silver-plated nylon) for a period of 36 h on each of the four subjects. The volunteers continued with their daily lives but were asked to return to the laboratory to perform measurements spaced at 0, 1, 3, 7, 12, 24, 30 and 36 h. The measurements obtained at time 0 were added to the dataset for the noise analysis. During the entire process, we did not remove the textile electrodes from the patient’s skin; nevertheless, we adjusted them against displacements on each partial measurement. Due to the duration of the experiment, we replaced the disposable electrodes on each measurement. We did not performed additional measurements like contact impedance during these tests.
Signal processing (like filtering) was the same described in the noise measurements section. As this study focuses on the performance of fabrics, no especial filtering or higher order filter was needed. The selected range of frequencies permits components to pass through and provides a high-fidelity tracing for the P–Q–R–S–T ECG wave. Consequentially, we used the segmentation algorithm, introduced above, to extract the P–Q–R–S–T complex from the ECG trace and split it into single P–Q–R–S–T waves for individual analysis. Each ECG segment from the textile electrodes was compared with the segments, captured simultaneously, from the Ag/AgCl electrodes; then, the error rate is calculated by dividing the number of complete ECG segments registered with the experimental material, against the number of ECG segments captured simultaneously with Ag/AgCl commercial electrodes. We analyzed the data through a multivariate ANOVA of repeated variables.
