Comprehensive Signal Quality Evaluation of a Wearable Textile ECG Garment: A Sex-Balanced Study

TL;DR

This study evaluates a new textile ECG garment designed to improve signal quality across sexes, demonstrating its effectiveness and highlighting the importance of sex-specific design for reliable wearable cardiac monitoring.

Contribution

Introduces a novel textile ECG device with optimized electrode placement and provides a comprehensive sex-balanced evaluation framework.

Findings

Achieves signal quality comparable to reference devices in rhythm and morphology.

Demonstrates robust classification performance for physiological parameters.

Identifies sex-specific factors influencing ECG signal acquisition.

Abstract

We introduce a novel wearable textile-garment featuring an innovative electrode placement aimed at minimizing noise and motion artifacts, thereby enhancing signal fidelity in Electrocardiography (ECG) recordings. We present a comprehensive, sex-balanced evaluation involving 15 healthy males and 15 healthy female participants to ensure the device's suitability across anatomical and physiological variations. The assessment framework encompasses distinct evaluation approaches: quantitative signal quality indices to objectively benchmark device performance; rhythm-based analyzes of physiological parameters such as heart rate and heart rate variability; machine learning classification tasks to assess application-relevant predictive utility; morphological analysis of ECG features including amplitude and interval parameters; and investigations of the effects of electrode projection angle given by the textile / body shape, with all analyzes stratified by sex to elucidate sex-specific influences. Evaluations were conducted across various activity phases representing real-world conditions. The results demonstrate that the textile system achieves signal quality highly concordant with reference devices in both rhythm and morphological analyses, exhibits robust classification performance, and enables identification of key sex-specific determinants affecting signal acquisition. These findings underscore the practical viability of textile-based ECG garments for physiological monitoring as well as psychophysiological state detection. Moreover, we identify the importance of incorporating sex-specific design considerations to ensure equitable and reliable cardiac diagnostics in wearable health technologies.

Tables

Table 1. TABLE I : Related work on textile-based ECG monitoring with the number of subjects, electrode setup, protocol, reference system, and selected quality assessment, sorted by date of publication. (F: Female, M: Male, U: Unknown)

Ref.	Study Population	Electrodes (el) Channels (ch.)	Protocol	Reference System	Quality Assessment
[39]	5 (5M), healthy	3 el. (RA, LA, LL), 2 ch. (mod. Einthoven)	standing, sitting (relaxed)	Biopac MP150 (Ag/AgCl)	R-Peak detection rate (estimated/observed)
\rowcolor[gray]0.9 [40]	40 (40M), cardiovascular diseases	2 (V5, V5R)	20 subjects (in bed), 20 subjects (mild exercise, cycloergometer, walking)	CL Delta / Fukuda Denshi DS 5700 (Ag/AgCl)	cardiologist rating (artifacts, morphological, rhythm), compute rule based rating
[41]	4 (4M), healthy	2 el. (chest belt), 1 ch. (custom)	standing, treadmill (4, 5, 6, 7 km/h)	Biopac MP150 (Ag/AgCl)	QRS delination, sensitivity, accuracy, TP, FP, FN
\rowcolor[gray]0.9 [42]	12* (12U), unknown	3 el. (reference + 2 custom), 1 ch. (custom)	rest and unknown motion	Synmedic Ambu Blue (Ag/AgCl)	subjective rating
[34]	10 (10U), healthy	various (task dependent)	sitting, breathing, stairs, 2Hz step-up-down-exercice	-	QRS detection rate, RMSE
\rowcolor[gray]0.9 [43]	3 (3M), unknown	9 el (chest), 8 ch. (12 Mason-Likar)	lying, sitting, walking	Corscience Holter ECG BT12 (Ag/AgCl)	Bland-Altmann, fiducial points, reference correlation
[44]	10 (10U), healthy	3 el (RA, RA, LL$†$ ), 1 ch. (mod. Lead I)	-	Unknown (Ag/AgCl)	R-peak amplitude and variations, SNR
\rowcolor[gray]0.9 [45]	4 (2M, 2F), healthy, minors	2 el. (chest), 1 ch. (custom)	resting, walking, running (12 km/h)	Cardiac Monitor Arm Band (Ag/AgCl)	visual inspection, HR, tachycardia/extra-systoles
[35]	1 (1U), healthy	3 el (LA, RA, LL), 1 ch. (mod. Lead I)	-	Analog Devices INA 2128 (Ag/AgCl)	visual inspection, reference signal correlation
\rowcolor[gray]0.9 [46]	1 (1F), healthy	3 el (chest band), 1 ch. (mod. Lead I)	sitting, walking	ADS1292ECG-FE (Ag/AgCl)	visual inspection, baseline wander, HF noise
[47]	4 (4U), healthy	3 el (LA, RA, LL), 1 ch. (mod. Lead I)	-	TMSI Proti7 (Ag/AgCl)	PSD, visual inspection, delineation
\rowcolor[gray]0.9 [36]	66 (47M, 19F$\ast$ ), healthy	2 el. (V5R, V5) 1 ch. (custom location)	supine rest, seated rest, trunk rotation, steps	Holter ECG Cardy 303 Pico+ (Ag/AgCl)	R-amplitude variations isoelectric line variations
[48]	1 (1F), healthy	3 el (LA, RA, LL), 1 ch. (Lead I)	resting	-	visual inspection, SNR
\rowcolor[gray]0.9 [38]	1 (1U), healthy	3 el. (LA, RA, LL), 1 ch. (mod. Lead I)	-	-	peak amplitude, SNR (PSD)
[49]	1 (1U), healthy	5 el. (LA, RA, LL, RL, V2), 4 ch. (mod. Einthoven, V2)	sitting, standing, walking, stair climbing	Texas Instruments ADS1298R (Ag/AgCl)	visual inspection
\rowcolor[gray]0.9 [50]	10 (10M), healthy	chest, unknown	relaxed, resting	-	PQRST delineation (recall, precision, F1)
[26]	1 (1U), healthy	3 el. (LA, RA, LL), 1 ch. (mod. Lead I)	resting	-	peak amplitude, pSQI, BasSQI
\rowcolor[gray]0.9 [51]	13 (6M, 7F), healthy	3 el. (LA, RA, LL, RL) 2 ch. (mod. Einthoven)	standing, treadmill, $n$-back task	Bittium Faros 360 NeXus-10 MKII Polar RS800 Multi SOMNOtouch NIBP	morphological, HR
[52]	30 (16M, 14F), healthy	2 el., 1 ch.	activity during daily living (unspecified)	3-lead holter monitor	subjective, R-R interval comparison
\rowcolor[gray]0.9 Ours	30 (15M, 15F), healthy	4 el. (LA, RA, LL, RL) 2 ch. (mod. Einthoven)	lying, sitting, walking, running	Bittium Faros 180 & BTL-08 Holter H600 (Ag/AgCl)

Table 2. TABLE II : SQIs for ECG signals used in this study with references and a short description.

Ref.	Name	Description
[66]	sSQI	Skewness of the signal
\rowcolor[gray]0.9 [66]	kSQI	Kurtosis of the signal
[67]	pSQI	Relative power in the QRS complex ${\int }_{5Hz}^{15Hz}P\left(f\right)df/{\int }_{5Hz}^{40Hz}P\left(f\right)df$
\rowcolor[gray]0.9 [69]	cSQI	Regularity of RR intervals
[68]	qSQI	Matching degree of R peak detection (Detectors: [70] and [71])
\rowcolor[gray]0.9 [51]	morphSQI	Morphological SQI based on weighted cardiac cycle median deviation
Ours	vmSQI	Mean absolute distance of complexes to the median complex.

Equations

$$\underset{\left\{\mathbf{R}|\genfrac{}{}{0ex}{}{{\mathbf{R}}^{-1}={\mathbf{R}}^{†}}{|\mathbf{R}|=1}\right\}}{\underbrace{\min }}\Vert \mathbf{AR}-\mathbf{B}{\Vert }_F^2$$

Full text

I Introduction

Electrocardiographic recordings serve as a fundamental diagnostic tool in modern medicine, providing invaluable noninvasive insights into the electrical activity of the heart and therefore the health of the cardiovascular system. Introduced by Willem Einthoven in the early 20th century, Electrocardiography (ECG)remains a cornerstone in clinical cardiology. Einthoven’s pioneering work laid the foundation for understanding the principles underlying ECG acquisition and interpretation [1, 2].

ECG signals are acquired through electrodes placed on the skin, capturing the electrical impulses generated by cardiac muscle de- and repolarization. These signals manifest as characteristic waves on the ECG tracing, reflecting the sequential activation of different regions of the heart [3].

In modern medicine, ECG is used in applications ranging from diagnosing cardiac arrhythmias [4] and ischemic heart disease [5] to monitoring patients during surgery [6] and assessing the effects of pharmacological interventions [7, 8]. Moreover, recent advancements in wearable devices have expanded ECG applications beyond traditional cardiology, enabling emerging research in affective sensing [9] and stress detection during daily activities [10].

Each of these applications have domain specific requirements regarding the quality of the ECG signal, the duration of the recording, usability, accessibility, the comfort of the patient and costs. While the traditional approach of using Ag/AgCl gel electrodes provides reliable signal quality, it also presents several limitations. These include the risk of skin irritation [11, 12], the limitation that only trained professionals are able to correctly apply the adhesive electrodes at specific anatomical locations [13], and restricted freedom of movement imposed on patients during daily activities by the presence of electrodes and connecting wires.

Recent advances in wearable technology have addressed specific challenges in health monitoring applications. Devices such as smartwatches [14] and smart rings [15] utilize Photoplethysmography (PPG) to measure heart rate (HR)and heart rate variability (HRV), thereby enabling applications such as activity tracking [16] and stress monitoring [17]. Despite their utility, these devices have limited accuracy in capturing the precise timing of heartbeats, as they measure the mechanical pulse of blood flow rather than the heart’s electrical activity. Consequently, their diagnostic capabilities are restricted, particularly for conditions such as ischemic heart disease or conduction disorders, as well as for obtaining accurate HRV parameter estimates, all of which require precise beat segmentation. Furthermore, their measurements are susceptible to noise arising from individual patient characteristics such as skin tone, weight, age, and gender as well as physiological factors including blood pressure and temperature, and external influences like ambient lighting conditions [18]. Another category of devices, such as the Apple Watch [19], is capable of recording brief segments of ECG waveforms and can automatically detect arrhythmias [20, 21]. Similarly, systems like AliveCor [22, 23] are utilized to assess morphological changes during pharmacological studies [24]. Smart textiles represent a unique class of wearable devices that can continuously record the heart’s electrical activity over extended periods [25]. Unlike conventional wearables, smart textiles offer enhanced comfort and freedom of movement, enabling unobtrusive and continuous ECG monitoring during daily activities. This is achieved by seamlessly integrating flexible dry electrodes into the garment itself, thereby further enhancing wearer comfort [26, 27, 28]. Various research groups and device manufacturers have developed different materials for this purpose, like metal based contact surfaces [29, 30, 31, 25, 32, 33], conductive polymers [34, 35, 25, 32, 33, 36] and carbon materials [37, 35, 38, 25].

These conductive materials are integrated into the textile using various fabrication techniques: Embroideries, knittings, sewings and weavings are used to classically integrate yarns [53, 54, 32, 55]. Others have printed [56], electrostatically flocked [57], or dip coated [37] conductive materials onto the textiles. Table I summarizes several studies that have introduced novel materials and integration techniques for textile electrodes. These efforts seek to optimize the signal-to-noise ratio, lower skin-electrode impedance, reduce motion artifacts through robust and stable skin-electrode contact, increase durability, including machine washability and maintain biocompatibility.

Significant differences exist in physiological parameters between male and female ECG recordings. For instance, women typically exhibit a higher HR, a lower ST-segment, and a reduced T-wave amplitude compared to men [58, 59]. Other anatomical factors, including variations in chest shape affecting electrode placement and differences in heart size [60], can result in ambiguous recordings, potentially leading to invalid diagnoses with certain electrode positions [61]. When developing a new system intended for use by both sexes, it is crucial to evaluate the system with a diverse study population. This ensures that the system is not biased and, as a result, can provide valid diagnostics [62, 63].

A primary motivation for developing new wearable sensor technology is to enable the continuous acquisition of high-quality ECG recordings. However, current research in this field reveals several limitations, including small sample sizes, study populations restricted to a single gender (typically male), and evaluation methodologies that focus exclusively on electrode materials rather than assessing the signal quality of the entire wearable system. These shortcomings, as detailed in Table I, give rise to two primary research questions: RQ1) How does a wearable ECG textile with integrated dry silicone electrodes perform in accurately measuring ECG parameters like HR, HRV, and morphology compared to a medical-grade gold-standard Holter ECG during different activities like sedentary tasks versus dynamic movements like walking or running? RQ2) To what degree does the structural design of the textile-based ECG system influence the quality of recorded signals, particularly in regard to addressing potential gender-related disparities in cardiac diagnostics?

II Methods

To evaluate the functionality and accuracy of our textile-based ECG device, participants simultaneously wore our wearable textile shirt with integrated dry electrodes, as well as two reference systems representing the current gold standard, each employing Ag/AgCl electrodes.

The textile incorporates four electrodes Left Arm (LA), Right Arm (RA), Left Leg (LL), and Right Leg (RL) providing three bipolar leads (two directly measured and one calculated), corresponding to modified Einthoven leads recorded at the torso. Figure 1 provides an overview of the garment with integrated electrodes and connections. The fourth electrode serves as the neutral electrode. Flexible electrodes entail a multilayer structure comprising $78\%$ polyamide and $22\%$ elastomer stretch-tricot, metal-plated with silver ($26.9\pm 1\%$), topped off by a conductive silicone sheet. This composition ensures durability and low surface resistance, while its stretch properties provide comfort during wear. The electrode cables are seamlessly integrated into the textile and routed to a central hub positioned at the back of the subject’s neck, where the connectors are located. These connectors feature buttons to which the hardware can be attached. The electronics are responsible for digitizing the signal using a ADS1293 (Texas Instruments, USA) analog front-end. The data storage on an SD card is handled by the microcontroller nRF52840 (Nordic Semiconductor, Norway). During setup, the live signal is monitored via Bluetooth transmission, while recorded data stored on the SD card is used for subsequent evaluation, ensuring data consistency without potential loss of wireless connections. As reference systems, we used two commercially available medical devices: the BTL (BTL GmbH, Dornstadt, Germany) BT-08 Holter monitor and the Bittium (Bittium Corp., Oulu, Finland) Faros 180. Both devices are approved for medical use and were operated with Ag/AgCl wet electrodes.

II-A Study Population

The study population consists of $30$ subjects, with $15$ males and $15$ females. The minimal age is $21$, the maximal age is $60$, and the mean age is $30.2$ with a standard deviation of $8.4$. The height distribution for males and females, as well as the weight distribution, is $181.7\pm 5.1\,\mathrm{cm}$, $168.2\pm 6.8\,\mathrm{cm}$ and $84.7\pm 11.3\,\mathrm{kg}$, $60.7\pm 7.4\,\mathrm{kg}$ respectively. We measured the under chest girth according to the ISO 8559 [64] standard. The lower chest girth is $93.9\pm 7.5\,\mathrm{cm}$ and $75.9\pm 4.6\,\mathrm{cm}$ for males and females respectively. All subjects participated voluntarily, had no known cardiovascular disease and were not taking any medication, except oral contraceptives, that could influence the ECG signal. The study was conducted in accordance with the Declaration of Helsinki and was approved by the Ethics Committee of Friedrich-Alexander-University Erlangen-Nuremberg (65_ 21 B, approved on 16.03.2021).

II-B Measurement Locations and Reference Systems

Since Einthoven introduced the three standard limb leads, several modifications have been made to improve the practicality of lead placement. For example, during treadmill ECG exercise, electrodes are attached to the back to reduce movement artifacts. In mobile Holter systems, the electrodes are usually placed on the torso in the direction of the standard Einthoven limb leads to allow for greater freedom of movement. In our reference systems, the upper electrodes were mounted along the anterior axillary line (over the bone) at the level of the clavicle, while the lower electrodes were placed outside the anterior axillary line at the end of the ninth costal cartilage. For the BTL five-lead system, the precordial electrode was positioned at the fourth intercostal space on the right margin of the sternum (lead V1). This electrode placement was selected because it is a common scheme for Holter monitors and provides sufficient space to accommodate the textile system.

The textile electrodes are positioned closer to the heart in alignment with the limb leads. The lower electrodes are placed at the fifth intercostal space on the anterior axillary line. These modifications are made to ensure a tight fit and minimize movement artifacts resulting from limb motion due to the greater distance. The garment was designed to ensure that the orientation of the electrodes closely aligns with natural physiological contours. Specifically, the upper electrodes were positioned at a 45-degree angle relative to the axillary line, while the lower electrodes were placed horizontally with a slight inward rotation shift towards the center. This configuration allows the electrodes to fit to the torso without protruding, thereby following the body’s shape for optimal contact and comfort. The garment is designed with adjustable hook-and-loop fasteners positioned at both the upper and lower regions, enabling precise adaptation to various body shapes and ensuring a tight fit. To accommodate more substantial variations in body size, multiple garment sizes were produced in accordance with the DIN EN 13402-3:2017-12 standard, ranging from XS to XXL. The actual fit of the garment and the positioning of the hook-and-loop fasteners varies according to individual body shape, which may, in turn, affect the placement of the electrodes. To systematically assess electrode positioning, we measured several key distances, as illustrated in Figure 3(a) and Figure 3(b). These figures depict the electrode locations for both the reference and textile systems, as well as the measured distances between the wearable electrodes and relevant anatomical landmarks. This approach enabled us to evaluate electrode positioning and analyze the impact on the anatomical projections.

II-C Study Protocol

The study protocol comprised a series of four-minute tasks, beginning with an initial seated resting phase, followed by an n-back working memory task. In the n-back task, participants were presented with a sequence of stimuli and asked to indicate whether the current stimulus matched the one presented n steps before, thereby inducing a state of high cognitive load [65]. Subsequently, participants completed a range of physical tasks presented in randomized order: standing motionless in an upright posture; walking at a moderate pace (up to $5\,\mathrm{km/h}$) with speed adjustments allowed for comfort; running at a brisker pace (up to $10\,\mathrm{km/h}$) with speed modified for individual fitness and body characteristics; cycling on a stationary bike at a target speed of $20\,\mathrm{km/h}$, with adjustments permitted; and lying down in four different positions right side, left side, back, and prone position each for four minutes. The protocol concluded with a final four-minute seated period. Throughout, procedures were tailored to ensure participant safety and comfort while preserving experimental rigor.

II-D Signal Quality Evaluation

In this study, we conduct signal quality evaluations without the use of human annotation or subjective feedback by domain experts. We assess the signal quality from multiple perspectives, each addressing distinct properties of the acquired signals. Signal quality measures, commonly employed by related literature, are extracted to compare devices during identical physical activity types. Additionally, physiological features based on rhythmic patterns, are extracted and compared between devices to further assess the integrity of the signals. We are accompanying this rhythm based evaluation by conducting a statistical evaluation of beat-detection performance. Morphological features, which hold significant clinical and pharmaceutical relevance, are also examined in this study. Specifically, we analyze features such as QT intervals, commonly assessed in pharmaceutical research, as well as amplitude markers associated with ventricular de- and repolarization. These amplitude markers are frequently utilized in clinical practice, for instance, in the detection of cardiac ischemia. Finally, we consider morphological changes that may intrinsically result from alternate electrode placement, acknowledging that such variations can influence the overall signal characteristics. This multifaceted evaluation strategy ensures a robust and objective assessment of signal quality. Initial preprocessing involves signal filtering, followed by segmentation procedures, including R-peak detection and fiducial point delineation. Detailed descriptions of these methods are provided in the Supplementary Material to ensure reproducibility while maintaining focus on the main research objectives.

II-D1 Signal Quality Indices

We utilize Signal Quality Indices (SQIs) to assess each acquisition device individually. These SQIs are then compared across devices and tasks. Previous studies [66, 67, 68, 51] introduced these indices for evaluating the quality of ECG signals. Table II presents the names, descriptions, and corresponding references for each Signal Quality Index (SQI). These indices capture multiple dimensions of signal quality, including morphological characteristics, spectral content, and measures of signal regularity.

We calculate the SQIs for all leads and compare them across different phases. Previous studies have attempted to establish thresholds based on heuristic assumptions or trained classifiers using expert annotations on multiple SQIs. However, because these heuristics and classifiers are typically tailored to specific tasks and rely on subjective ECG quality assessments, we calculate SQIs to enable an objective comparison between different acquisition devices.

II-D2 Rhythmic Based Evaluation

The second approach centers on the extraction and evaluation of widely recognized features associated with rhythmic physiological changes, specifically HR and HRV. In line with our initial methodology, we systematically compare these features across multiple acquisition devices, focusing on core parameters such as HR, the Root Mean Square of Successive Differences (RMSSD), the Standard Deviation of RR-Intervals (SDNN), and the High Frequency (HF) and Low Frequency (LF) spectral components. These metrics are extensively utilized in clinical contexts to assess autonomic function and cardiovascular health [72]. Given that the accurate estimation of these parameters depends on the reliable detection of heartbeats, typically achieved by identifying R-peaks in the ECG signal, we conduct a focused evaluation of R-peak detection performance. Specifically, we evaluate detection quality across three distinct devices by selecting ECG channel positions analogous to lead II, which is commonly employed in peak detection tasks [73]. The classification of detection events is illustrated in Figure 4, which provides a visual summary of typical signal artifacts and detection errors encountered during an active recording session. The figure highlights various categories of errors, including missed detections and invalid detections across different device channels. Notably, the depicted interval exhibits considerable movement and muscle noise artifacts.

To comprehensively assess detection performance, we compare a reference device with the textile-based system across both male and female participants and throughout all experimental phases, accompanied by evaluating the reference devices with each other. Both the BTL and Faros systems utilize Ag/AgCl electrodes positioned in proximity on the body, thereby providing a robust reference measurement. To further elucidate the influence of signal projection, we report hit, miss, and false alarm rates of the lead I and V1 to the reference system Lead II. This analysis underscores the impact of electrode positioning and channel selection on detection fidelity. To quantitatively compare the raw detection rates observed across different devices, we estimate the parameters of a Dirichlet distribution using Maximum Likelihood Estimation (MLE) [75]. Probabilistic modeling of detection events enables the performance of different devices to be represented by the parameter vector $\boldsymbol{\alpha}$, facilitating a robust comparison between detection events of different devices.

II-D3 Classification

We trained classifiers on surrogate tasks to assess the predictive performance for common ECG applications using different acquisition devices. To accomplish this, we introduce two commonly evaluated tasks in ECG signal processing and train classifiers on the ECG signals. Both tasks employ rhythmic features, consistent with established applications in physiological signal processing of ECG signals. The HRV features include time-domain measures, such as the SDNN and the Standard Deviation of the Successive Differences (SDSD). Frequency-domain features are also utilized, specifically the low-frequency (0.04–0.15 Hz) and high-frequency (0.15–0.4 Hz) components, as well as the ratio of high-to-low frequency spectral components (HF/LF). In addition, features derived from the graphical Poincaré plot the SD1 and SD2 components and their ratio (SD1/SD2) are incorporated. All features are extracted from lead II. Our activity recognition task is formulated as a three-class problem using a balanced dataset comprising the classes Active, Sitting, and Lying. The Active class includes activities performed during cycloergometer use, treadmill running, and treadmill walking. The Sitting class encompasses periods when the subject is sitting relaxed, watching a video, or participating in the n-back experiment. The Lying class captures instances where the subject is lying on their back, left side, or in prone position. The second machine learning task utilizes the same feature set as input, but is designed to distinguish between two classes reflecting different psycho-physiological states. The Low load class represents a condition where the subject is sitting upright at rest without any imposed task. In contrast, the High load class corresponds to the state in which the subject is engaged in solving the second level of the n-back test, a condition known to induce a high level of task load [65, 9]. We trained our models using lead II data from both the primary reference system and the textile-based system. Feature extraction was performed for each phase, followed by normalization via z-score transformation to achieve zero mean and unit standard deviation. For classification, we employed XGBoost, a model widely used in feature-based detection of psycho-physiological states [9]. The dataset was partitioned on a subject-wise basis using a 5x5 nested cross-validation scheme. We report the mean and standard error of the Area Under the Curve (AUC) and $F_{1}$ score for both tasks and devices.

II-D4 Morphological Comparison

The morphological comparison of ECG recordings obtained simultaneously from two devices, each employing distinct electrode placements, presents inherent challenges due to the differing spatial projections resulting from the cardiac electrical activity. Specifically, the orientation and placement of electrodes significantly affect the measured signals, as each device records the cardiac electrical field from a distinct spatial perspective. Integrating electrodes into a textile demands further adjustment of lead positions to fulfill both ergonomic requirements and the minimization of noise artifacts due to relative electrode-skin movement, as depicted in Figure 2. In addition, the use of the aforementioned silicon-based electrode material, in contrast to the Ag/AgCl wet electrodes used in the reference system, exhibits distinct characteristics with respect to modeling the skin–electrode interface [76, 77]. These modifications are essential to ensure user comfort and robust signal acquisition but, in turn, complicate direct morphological comparisons between signals derived from disparate lead configurations. Non-standard Einthoven placement of electrodes without detailed indication might cause diagnostic issues [78].

As an initial step in our morphological assessment, we extract commonly used fiducial points in ECG analysis and compute both interval- and amplitude-related measures. Specifically, we determine the amplitude from the isoelectric line to the R wave, the amplitude of the ST segment, the amplitude from the isoelectric line to the T wave, and the amplitude from the isoelectric line to the P wave. Additionally, we calculate the QT interval, representing the time duration from the onset of the Q wave to the end of the T wave, as well as the T-length, defined as the interval between T-onset and T-offset (extracted using tangents [79]), as shown in Figure 5. Direct projection or transformation of ECG signals between such configurations is nontrivial, as the inverse problem is ill-posed and affected by factors such as respiratory motion, the variable composition of intervening tissues (e.g., skin and fat layers), and, critically, the relative orientation of each electrode set with respect to the cardiac vector [80]. Recognizing these complexities, we approximate the correspondence between the two electrode systems as a transformation comprising rotation and linear scaling within the lead I/II plane. To formalize this, we frame the problem as a one-sided Procrustes analysis, where the objective is to optimally align the median complexes extracted from the textile-based recordings, denoted as $\mathbf{A}\in\mathbb{R}^{d\times n}$, to those from the reference BTL system, denoted as $\mathbf{B}\in\mathbb{R}^{d\times n}$, with $d=2$ corresponding to leads I and II and $n$ to the median complex sequence length. The transformation is modeled as a right-sided operation, with parameters determined by minimizing the squared Frobenius norm $\|\cdot\|_{F}^{2}$, as shown in Equation 1.

[TABLE]

To enhance alignment, we first normalize the textile-based median complex $\mathbf{A}$ to match the maximal amplitude observed in the reference system $\mathbf{B}$. Specifically, for each sample in both $\mathbf{A}$ and $\mathbf{B}$, we compute the instantaneous magnitude across leads I and II as $|\mathbf{v}|=\sqrt{I^{2}+II^{2}}$. We then determine the normalization constant $s$ as the ratio of the maximum absolute magnitude in the reference system to that of the textile system, that is, $s=\max(|\mathbf{v}_{B})|/|\max(\mathbf{v}_{A}|)$. The textile signals are subsequently scaled by this constant prior to transformation, ensuring that both datasets are on a comparable amplitude scale and thereby improving the robustness of the subsequent Procrustes alignment. To quantitatively assess the quality of alignment between the transformed textile-based and reference ECG signals, we employ three similarity measures: the squared Frobenius norm error, the Root Mean Square Deviation (RMSD), and the cosine similarity. Upon determination of the optimal rotation parameters within the I/II plane, we further correlate these findings with individual anthropometric measurements to elucidate which physiological attributes most significantly influence the effects of altered electrode placement. This integrative approach not only enables rigorous morphological comparison between textile and reference ECG recordings but also advances our understanding of the anthropometric determinants impacting signal morphology under novel electrode configurations.

III Results

III-A Comfort and Fit

The subjects were provided with appropriately sized garments, with size distribution as follows: among male participants, none wore size XS or XXL, while one wore size S, four wore size M, six wore size L, and four wore size XL. Among female participants, one wore size XS, ten wore size S, four wore size M, and none wore sizes L, XL, or XXL. The wearable garments were designed with a unisex sizing scheme. Participants were subsequently asked to assess their subjective comfort using a five-point Likert scale, ranging from 1 (very uncomfortable, barely tolerable, and highly disturbing) to 5 (very comfortable, such that the wearer forgot they were wearing the textile). The overall comfort ratings for male participants were distributed as follows: two rated the textiles as 2, six as 3, five as 4, and two as 5. For female participants, eight assigned a rating of 3, six a rating of 4, and one a rating of 5.

III-B Signal Quality

III-B1 SQIs

For cSQI, the lowest values are observed in BTL, Faros, and Textile Lead II. Other leads not only exhibit higher values during lying and sitting phases but also increase during more rigorous activities, such as treadmill running. Given that cSQI is primarily based on RR interval detection, these findings suggest that the projection of Lead II is optimal for R-peak detection, with device-specific differences playing a minor role in this particular SQI. The ksQI highlights the pronounced effects of intense activities, with the lowest values recorded during cycloergometer and treadmill running phases. morphSQI reveals the most pronounced changes across all SQIs for the textile electrode. This observation underscores that the morphology of individual heart complexes can differ from the surrounding signal, potentially due to baseline wander or other low-frequency variations. Regarding pSQI, although the median values remain similar across all phases, leads I and V1 exhibit greater variance compared to lead II. This further emphasizes the impact of lead placement, which alters the ratio between low and high frequency components in the QRS complexes. The qSQI demonstrates lower values during lying phases, while sitting and standing phases yield the highest values. This index reflects the degree of concordance between two R-peak detectors, while lead II of all systems generally has higher values compared to lead I and lead V1. sSQI shows that leads I and V1 consistently yield lower values than lead II across all devices. For vmSQI, variance and changes are generally minimal across all leads, except during running and walking phases, where increased variance is observed. In summary, while several parameters suggest that lead II performs comparably to the reference systems even at varying activity intensities, the statistical evaluation alone does not provide a comprehensive assessment. Lead placement exerts a more significant influence on the computed SQIs than artifacts such as baseline wander, muscle noise, or movement artifacts. Therefore, a thorough evaluation should prioritize practical applicability and intended use rather than relying solely on statistical features.

III-B2 Physiological Parameters

Mean HR, as well as HRV indices RMSSD and SDNN, together with the LF/HF spectral ratio, are widely employed in fitness and psychophysiological assessments. To evaluate the agreement between devices, we present these parameters using Bland-Altman plots, as illustrated in Figure 7. The results for HR and time-domain HRV measures show close agreement between devices for the majority of measurements. The observed outliers are typically attributable to errors in R-peak detection. Specifically, the mean difference is -0.38 for HR, -0.93 for RMSSD, -0.86 for SDNN, and -0.06 for the LF/HF ratio, indicating a high level of concordance between the measurement devices.

No systematic errors are observed, such as higher values resulting in greater deviations between the two devices. Visualizations in our supplementary material, reveal a high correlation of up to $R=0.93$, with minor reductions in correlation primarily attributable to R-peak detection errors. We present the hit rate, miss rate, and false alarm rate for all leads, compared to the reference device, across all phases in Table III-B2, separated by sex. The results indicate that the textile performs equally well for both female and male subjects, achieving a hit rate of $0.95\pm 0.10$ for females and $0.95\pm 0.12$ for males. This performance is comparable to that of the secondary reference device, Faros, which demonstrates hit rates of $0.96\pm 0.07$ for females and $0.97\pm 0.04$ for males. Additionally, we report the R-peak detection performance for lead I, showing lower performance for females compared to males, indicating a strong impact of RA/LA electrode placement on R-peak detectability.