Self‐Maintaining Electrode–Skin Interface for Motion‐Robust Biosignal Monitoring

ABSTRACT

The electrode–skin interface presents a fundamental challenge in bioelectronics: maintaining stable electrical contact with tissue that undergoes constant physiological changes. Here we demonstrate a paradigm shift by engineering a self‐sustaining microenvironment at the skin–electrode interface through the integration of Au‐coated fabric electrodes with a humidity‐regulating cooling patch. Our smart electrode‐integrated cooling (SEIC) patch uses evaporative cooling to create a humid microenvironment where water condensation continuously regenerates ionic pathways between skin and electrode. This transforms the interface from a static contact point to a dynamic, self‐renewing electrochemical junction. The SEIC patch exhibits 200‐fold and 10‐fold lower impedances than Au‐coated fabric electrodes and conventional electrodes, respectively; maintains performance through 2000 attachment cycles; enables stable biosignal acquisition for eight days; and preserves signal fidelity under motion. Moreover, the SEIC demonstrates applications in continuous cardiac monitoring during daily life, and real‐time 3D facial animation reconstruction in virtual reality via high‐fidelity capture of facial muscle activity. This work opens new frontiers in preventive medicine and human–computer interaction.

A smart electrode‐integrated cooling patch creates a humid microenvironment that continuously regenerates ionic pathways between skin and electrode, transforming the interface into a dynamic, self‐renewing electrochemical junction. This enables stable biosignal acquisition under motion and allows continuous cardiac monitoring and real‐time 3D facial animation reconstruction.

1. Introduction

Wearable electrodes that non‐invasively interface with human skin to record biopotential signals from organs offer significant advantages for healthcare monitoring, human–computer interaction, and rehabilitation [1, 2, 3, 4, 5, 6, 7]. However, one of the fundamental challenges in bioelectronic sensing lies at the electrode‐skin interface: establishing stable and reliable electrical contact with skin that is constantly undergoing physiological changes. This interface must consistently convert ionic currents from the body into electronic currents for the device, while simultaneously accommodating mechanical deformation, variations in temperature and humidity, changes in skin adhesion, and the natural exfoliation of the stratum corneum [8, 9]. These interface challenges manifest as signal‐to‐noise ratio (SNR) degradation across multiple timescales‐from instantaneous impedance spikes caused by motion to gradual baseline drift over several hours. The effects are most clinically pronounced during physical movement, when mechanical coupling between the electrode and skin is repeatedly disrupted [10, 11]. Each body movement induces microscopic separations and recontacts at the interface, generating electrical artifacts that can resemble pathological signals. This fundamental mismatch between static electronic components and dynamic biological surfaces remains an unresolved issue, limiting the long‐term reliability of biosensing technologies in applications ranging from intensive care cardiac monitoring to next‐generation human–computer interface [12, 13, 14, 15].

These interface challenges are perhaps most consequential in cardiac monitoring, where motion artifacts in wearable electrocardiography (ECG) systems can have life‐threatening implications. ECG represents the most mature and widely deployed biosensing modality, yet even after decades of clinical use, the fundamental problem of motion‐induced signal degradation remains largely unsolved. Unlike other biosignals where artifacts merely reduce data quality, ECG artifacts can mimic life‐threatening arrhythmias, creating a dangerous ambiguity between noise and genuine cardiac events. For example, in a study of 1000 ambulatory ECG recordings, motion artifacts led to a 3.5% misinterpretation rate with potentially life‐threatening consequences [16]. Distinguishing artifacts from true arrhythmias is difficult‐even among trained professionals‐leading clinicians to discard 20–40% of data during stress tests to avoid misdiagnosis [17, 18, 19]. This creates dangerous blind spots, especially since exercise‐related cardiac arrests, though rare (2.2%), have much higher survival rates (57.7%) compared to non‐exercise cases (12.6%) [20].

Despite decades of innovation, the unstable electrode–skin interface remains the root cause of signal degradation, producing impedance fluctuations of 20–80% during everyday activities due to changes in skin hydration, temperature, and sweat gland activity [8, 21]. The persistence of these clinical challenges stems from a fundamental mismatch: while hardware improvements (e.g., noise‐rejecting amplifiers [22], active electrodes [23], shielded cables [24]) and algorithmic approaches (adaptive filters [25], machine learning [26]) have advanced significantly, they still operate on compromised data. In real‐world settings, machine learning models often underperform due to the variability of interface conditions, falling from > 99% accuracy in labs to as low as 60–75% [27, 28]. Without a stable and consistent electrochemical interface, even advanced signal processing cannot reliably isolate true biological signals from noise [29].

Recognizing this fundamental limitation, researchers have pursued various strategies to improve the electrode‐skin interface itself through material innovations. However, the challenge of electrode design extends beyond optimizing electrical performance to include practical requirements that influence commercial viability and clinical adoption. Conventional gel electrodes, although offering low initial impedance via an electrolyte bridge, have significant limitations: adhesion typically degrades within 4–24 h due to gel dehydration [5], skin irritation affects 5–15% of users‐limiting long‐term wearability [2], and single‐use designs create substantial cost and waste, making them unsuitable for continuous monitoring. As a result, healthcare facilities increasingly seek reusable electrodes capable of withstanding hundreds of uses for both economic and environmental reasons [30]. Recently, an increasing number of researchers have focused on advancing wearable electrode technologies. One area of research is dry thin‐film such as metal electrodes [4, 14, 31, 32], conductive polymers [1, 3, 33], and conductive textile‐based electrodes [34, 35, 36, 37]. Another direction involves conductive hydrogel [38, 39, 40, 41], which are valued for their softness and biocompatibility, making them suitable for prolonged skin contact. A third category is tattoo‐based electrodes [42, 43, 44, 45, 46], which are ultrathin and form tight, conformal contact with the skin, minimizing air gaps and significantly reducing skin‐electrode impedance. A fourth emerging category is 2D materials [47, 48, 49, 50], nanoscale skin‐conformal electrodes that improve durability and mechanical compliance, enabling continuous resilient electrophysiological signal monitoring with highly suppressed noise, low motion artifacts, and high‐water resilience [49, 50]. Although these electrodes have shown considerable advancements, several challenges remain. For example, these materials still struggle to maintain consistent performance due to air gaps, moisture, temperature fluctuations, poor adhesion to the metal transition layer, limited reusability, and mechanical stress during daily use [2, 46, 51]. Ultimately, maintaining stable electrical coupling under dynamic environmental conditions remains the core challenge.

In this work, we present a fundamentally different approach by engineering a self‐sustaining microenvironment at the skin–electrode interface through the integration of Au‐coated fabric electrodes with a humidity‐regulating cooling patch. Rather than forcing stable contact through materials alone, our smart electrode‐integrated cooling (SEIC) patch creates controlled conditions that actively maintain electrical coupling. Through evaporative cooling, the patch establishes a humid microenvironment where microscopic water condensation continuously regenerates the ionic pathways between skin and electrode. This controlled condensation process—similar to how morning dew reliably forms on cooled surfaces—provides a constant source of ionic conductivity at the interface. This transforms the interface from a static contact point to a dynamic, self‐renewing electrochemical junction.

This microenvironment engineering delivers transformative improvements across multiple scales of validation, from fundamental electrical performance to clinical applications. At the materials level, the SEIC patch demonstrates 200‐fold lower impedance than Au‐coated fabric electrodes and 10‐fold lower impedance than commercial Ag/AgCl electrodes. The patch maintained functionality through 2,000 attachment‐detachment cycles, addressing the reusability challenge that plagues current electrodes. At the physiological level, the SEIC patch's high signal acquisition capability was validated by the SNR during the recording of ECG, electrooculography (EOG), and electromyography (EMG) signals, even in the presence of motion artifacts. It enabled continuous ECG monitoring for eight days, through various daily activities, without signal degradation. In addition, the SEIC patches were tested in a 3‐lead configuration with ten participants to measure ECG signals. The results demonstrated consistent acquisition of high‐accuracy signals with strong F1‐scores under diverse conditions, including sitting, walking, running, and cycling.

In real‐world deployment, SEIC patches were used to develop a wearable ECG device capable of continuous daily monitoring across thirteen typical daily activities. These findings highlight broader application domains. For example, cardiac patients could benefit from uninterrupted monitoring during exercise when the risk of sudden cardiac arrest is elevated. Stroke rehabilitation could be quantified through real‐time biosignal tracking during therapeutic movements. Furthermore, we integrated the SEIC patch into a VR headset for facial biopotential acquisition. By capturing facial muscle signals in real time and applying machine learning, we reconstructed 3D facial expressions with high fidelity. This system not only tracked emotional expressions dynamically but also provided proof‐of‐concept for human–computer interaction (HCI) platforms where subtle physiological states can be mapped onto immersive environments. Beyond HCI, 3D facial‐expression reconstruction also offers valuable healthcare monitoring capabilities, including early detection of neurological impairments, objective pain or stress assessment, and improved remote evaluation of facial‐motor function.

This work represents a paradigm shift in how we approach the fundamental challenge of interfacing electronics with biological systems. Rather than viewing the skin‐electrode junction as a problematic boundary to overcome, we demonstrate that engineering the microenvironment at this interface can transform it into a robust, self‐maintaining system. Our work not only tackles long‐standing challenges like motion artifacts but opens new frontiers in continuous health monitoring and human‐computer interaction.

2. Results and Discussions

2.1. Design and Functionality of the SEIC Patch

To create a self‐maintaining microenvironment at the skin–electrode interface, we design SEIC patch to harness evaporative cooling to create conditions for continuous ionic pathway regeneration. Rather than relying on static materials to maintain contact, our approach creates a dynamic interface that actively sustains the conditions necessary for stable bioelectronic coupling.

The structure of the SEIC patch includes: (i) a gold‐coated fabric electrode (8 mm diameter) soldered to a connection wire using silver epoxy; (ii) a 20 × 20 mm hydrogel temperature‐regulating cooling patch with a central hole through which the wire passes to integrate with the fabric; and (iii) a moisture‐resistant layer surrounding the hydrogel cooling patch to reduce vapor evaporation as shown in Figure 1a. More details on fabrication process of SEIC is described in Figure S1. The Au‐coated fabric is highly conductive, flexible, and moisture‐retainable, while the hydrogel temperature‐regulating cooling patch offers high water content, stable temperature, self‐adhesion, conformability, and biocompatibility. By integrating these two components, the SEIC patch creates the foundation for a self‐maintaining biointerface. When attached to human skin, it provides evaporative cooling to lower skin temperature and creates a localized, high‐humidity microenvironment, a thin water layer from condensation, at the skin–Au electrode interface (Figure 1b). This water layer maintains ionic pathways between tissue and electrode, creating a stable electrochemical interface that resists degradation over time (a more detailed discussion is presented in the section on fundamental characteristics of the SEIC patch). Moreover, the SEIC patch offers conformability, biocompatibility, and self‐adhesion without additional adhesives. Overall, this configuration enables stable, high‐quality biopotential recordings both at rest and during motion (Figure 1c).

FIGURE 1.

Design and functionality of the SEIC patch. (a) Schematic structure of the SEIC patch. (b) Illustration of the SEIC patch on human skin, showing the formation of a high‐humidity microenvironment at the skin–electrode interface. (c) Key features of the SEIC patch include high conductivity, humidity and temperature stability, low skin–electrode impedance, good adhesion, conformability, and biocompatibility, enabling a high SNR for biopotential signal recording‐even during motion artifacts. (d) Schematic of the wearable ECG device connected to an OpenBCI (brain‐computer interface) system on the human body. (e) Typical ECG waveform with labeled PQRST (the characteristic waveform components of an electrocardiogram: P wave, QRS complex, and T wave) intervals. (f) ECG signals recorded by the wearable ECG device during daily activity. (g) Integration of the SEIC patch into the facial interface of a VR headset. (h) Recording of EOG and EMG signals from facial expressions, followed by machine learning‐based signal processing to reconstruct 3D facial animations of the VR user.

We demonstrate SEIC's clinical potential and versatility through two main applications: (1) a motion‐resilient and continuous ECG sensor integrated on a body‐worn (Figure 1d) to monitor cardiac signals during daily activities (Figure 1e,f) and (2) a highly sensitive VR headset that supports continuous and real‐time 3D facial animation based on biosignals (e.g., EOG and EMG) captured on the face (Figure 1g,h).

2.2. Fundamental Characteristics of the SEIC Patch

To understand how the SEIC patch maintains stable bioelectronic interfacing, we characterized its fundamental components, thermal–moisture regulation capabilities, and electrochemical performance of the SEIC patch, highlighting how this self‐sustaining interface enables unprecedented impedance stability, reusability, and longevity.

The self‐maintaining microenvironment of SEIC is the result of two integrated components: The Au‐coated fabric, with a low sheet resistance of 0.3 Ω □⁻¹, provides the electrode component while its inter‐fiber spacing texture, as shown in Figure 2a,b, allows it to effectively absorb and retain moisture. The hydrogel temperature‐regulating cooling patch serves as a self‐adhesive substrate due to its ability to maintain a stable temperature, high water content, biocompatibility, and strong adhesion to human skin (Figure 2c). By integrating the Au‐coated fabric with the hydrogel cooling patch (Figure 2d), the SEIC patch creates conditions where water vapor continuously condenses at the cooled interface, maintaining the ionic pathways even during mechanical deformation. The SEIC patch achieves stable bioelectronic interfacing through a self‐maintaining microenvironment created by evaporative cooling.

FIGURE 2.

Fundamental characteristics of SEIC patch. Field emission scanning electron microscopy (FE‐SEM) images of Au‐coated fabric at low (a) and high (b) magnification. (c) Optical image of the hydrogel cooling patch showing strong adhesion to human skin. (d) Optical image of Au‐coated fabric integrated with the hydrogel cooling patch. (e) Skin temperature over 24 h with the Au‐coated textile and the SEIC patch applied. (f) Skin moisture over 24 h with the Au‐coated fabric and the SEIC patch applied. (g) Optical image of SEIC patches applied on the forearm. (h) Skin‐electrode contact impedance of the Au‐coated fabric electrode and the SEIC patch at frequencies of 200 Hz, 512 Hz, and 1022 Hz. (i) Skin‐electrode contact impedance of the Au‐coated fabric electrode and the SEIC patch at 512 Hz over a 24 h period. (j) Skin‐electrode contact impedance of 3 M, Kendall, and SEIC patches at 200 Hz, 512 Hz, and 1022 Hz. (k) Skin‐electrode contact impedance of 3 M, Kendall, and SEIC patches at 512 Hz under repeated attach–detach cycles.

The key innovations of the proposed mechanism are the following. The Au‐coated fabric integrated with the hydrogel cooling patch absorbs and retains moisture, creating a wet electrode surface which facilitates ion transport and improves electrode‐skin coupling. When the SEIC patch is attached to human skin, it provides evaporative cooling to lower skin temperature and creates a localized, high‐humidity microenvironment (a thin water layer from condensation) at the skin/electrode interface. This condensed water layer serves two purposes: (a) it hydrates the stratum corneum, reducing skin resistance, and (b) it combines with natural skin ions to form a temporary ion‐conductive film (assume as electrolyte layer), which enhances conversion from ionic current (in skin) to electronic current (in electrode) through an electrochemical interface. In contrast, standalone Au‐coated fabric electrodes lack this water layer. When attached to skin, they contact the dry stratum corneum through an air gap, resulting in poor adhesion and high interfacial resistance that degrades signal quality. This elimination of the air gap and formation of a continuous ionic pathway explains the dramatic impedance reduction observed in our measurements. The electrical equivalent models of the electrode–thin water layer–skin interface (for the SEIC patch) and the electrode‐skin interface (for the Au‐coated electrode) are presented and discussed in detail in Figure S2, to visualize and compare the skin–electrode contact impedance of the SEIC patch and the standalone Au‐coated fabric on human skin.

The following measurements confirm this mechanism and quantify its impact on electrode performance. To quantify the microenvironment's stability, we measured the temperature and moisture content of both the standalone Au‐coated fabric and the SEIC patch (Au‐coated fabric integrated with the hydrogel temperature‐regulating cooling patch) for comparison. The temperature of the SEIC patch was consistently lower than that of the standalone Au‐coated fabric (Figure S3a). Importantly, the SEIC patch exhibited significantly higher moisture content compared to the Au‐coated fabric, as shown in Figure S3b. When the hydrogel cooling patch contacts skin, it reduces the local temperature by 6°C (from 34°C to 28°C, Figure S3c) while increasing humidity from 40% to 89% (Figure S3d). This temperature differential drives continuous water vapor condensation at the electrode‐skin interface, forming a thin ionic conductive layer that bridges the electronic‐to‐ionic current conversion gap—the fundamental challenge in bioelectronic. The SEIC patch also maintained these conditions for extended periods: As shown in Figure 2e,f, skin with the SEIC patch maintained stable temperature and moisture compared to skin with only the Au‐coated fabric. Notably, skin with the SEIC patch was able to sustain stable temperature and moisture for up to 24 h.

These sustained microenvironmental conditions directly translate to superior electrical performance by reducing skin–electrode contact impedance. To demonstrate, we attached SEIC patches to the forearm (Figure 2g) and measured the skin–electrode contact impedance across a frequency range from 200 Hz to 100 kHz. As shown in Figure 2h and Figure S4, the SEIC patch exhibited a 200‐fold reduction in contact impedance compared to the Au‐coated fabric electrode. Specifically, the impedance values for the SEIC patch at 200, 512, and 1022 Hz were approximately 14, 6.4, and 2.9 kΩ, respectively. In contrast, the Au‐coated fabric electrode showed impedance values of approximately 3166 kΩ (200 Hz), 1488 kΩ (512 Hz), and 894 kΩ (1022 Hz). We also see that this dramatic impedance reduction persists over time. We evaluated the long‐term on‐skin impedance stability of the SEIC patch over 24 h and compared with the Au‐coated fabric electrode. Participants were allowed to continue normal daily activities, including exercise and showering. As shown in Figure 2i, Table S1, and Figure S5, the SEIC patch maintained stable and low skin–electrode contact impedance for up to 24 h. In contrast, the Au‐coated fabric electrode exhibited consistently high and unstable impedance throughout the same period. To further decouple the effects of temperature and humidity, we performed an additional controlled experiment under two conditions: (i) temperature‐controlled cooling without external humidity increase (using a cold patch at 24°C), and (ii) humidity‐enhanced condition achieved by pre‐wetted electrodes (≈85% relative humidity). As shown in Figure S6, both conditions reduce skin–electrode impedance compared to the non‐controlled Au‐coated fabric electrode; however, the impedance reduction is more pronounced under the humidity‐enhanced condition, while the temperature‐controlled condition alone also produces a measurable decrease. Importantly, the SEIC configuration still shows the lowest impedance among all cases, indicating a synergistic effect of cooling‐induced moisture accumulation and interfacial hydration.

In addition to performance gains, practical electrode systems must maintain their properties through extended use and repeated application‐removal cycles. To benchmark against existing technologies, the skin‐electrode contact impedance of the SEIC patch was compared with commercial electrodes: 3 M (dry conductive gel electrode) and Kendall (adhesive conductive gel electrode). The results in Figure 2j and Figure S7 show that the SEIC patch exhibits 10‐fold lower impedance than 3 M and sixfold lower than Kendall. To demonstrate reusability of the SEIC patch, we measured the skin‐electrode contact impedance after repeated attaching and detaching on human skin. As shown in Figure 2k, Table S2 and Figure S8, the SEIC patch maintained stable and low contact impedance even after 2,000 attach‐detach cycles. In contrast, the 3 M and Kendall electrodes showed a continuous increase in contact impedance with repeated use and completely lost skin adhesion after 32 and 400 cycles, respectively.

2.3. ECG Acquisition and Daily ECG Monitoring

We next evaluated the SEIC patch's performance in dynamic ECG monitoring scenarios to demonstrate its clinical utility. We tested both single‐lead and 3‐lead ECG configurations across multiple users and activity conditions, examining signal integrity, motion‐induced baseline drift, and signal‐to‐noise ratios.

The SEIC patches were attached to the subject's forearms and connected to an OpenBCI device, a portable wireless module with Bluetooth capability, to collect ECG signals in real time (Figure 3a). The captured ECG signals were compared with those measured using Au‐coated fabric, 3 M, and Kendall electrodes (Figure 3b). Characteristic ECG waveforms, including QRS complexes, P waves, and T waves, were successfully recorded. The SEIC patch achieved a SNR of approximately 13.5 dB, which was higher than that of the Au‐coated fabric electrode (≈11 dB) and comparable to the SNRs of the 3 M (≈13.1 dB) and Kendall (≈13 dB) electrodes (Figure 3c). Furthermore, long‐term ECG monitoring was conducted by attaching the SEIC patch to the subject's forearms for 8 consecutive days, with ECG signals recorded daily at 9 a.m. for 1 min. During this period, the subject engaged in various daily activities, such as walking, eating, cooking, working, showering, sleeping, and conducting laboratory experiments. We evaluated SEIC against state‐of‐the‐art ECG electrodes, including 3 M and Kendall electrodes. The SEIC patch maintained stable ECG waveforms and amplitudes throughout the 8‐day period (Figure S9). The SNR remained consistent and comparable to that of the 3 M and Kendall electrodes, with an average value of approximately 13 dB over the entire measurement period (Figure 3d). Moreover, the skin irritation was also investigated by monitoring the skin areas where SEIC, 3 M, and Kendall electrodes were attached on the forearm during ECG measurements over a 8‐day period. Optical images of the skin after electrode removal are provided in Figure 10. The results show that the skin in contact with the SEIC electrode remained comparable to normal skin, while mild redness was observed in areas where 3 M and Kendall electrodes were applied. These observations suggest that the SEIC electrode does not cause noticeable skin irritation under the tested conditions. In addition, due to its soft, flexible, and self‐adhesive design that does not require adhesive gels, the SEIC electrode minimizes potential sensory discomfort during wear.

FIGURE 3.

ECG monitoring during motion and daily activities. (a) Optical image of SEIC patches attached to the subject's forearms for capturing ECG signal. (b) ECG signals captured using SEIC patch, Au‐coated fabric, 3 M, and Kendall electrodes. (c) SNR of SEIC patch, Au‐coated fabric, 3 M, and Kendall electrodes during ECG acquisition. (d) SNR of SEIC, 3 M, and Kendall electrodes over 8 consecutive days of ECG monitoring. (e) SNR comparison of SEIC, Au‐coated fabric, 3 M, and Kendall electrodes during stationary and motion conditions. (f) Optical image showing the 3‐lead ECG configuration on the subject. (g) ECG signals (Lead I and Lead II) captured by SEIC patch in the 3‐lead configuration under various activities: sitting, walking, running, stair climbing, and cycling. (h) F1‐scores of SEIC, 3 M, and Kendall electrodes in a 3‐lead configuration for ECG detection during various activities. (i) Distribution of F1‐scores and accuracy across ten participants using SEIC patches in a 3‐lead configuration. (j) Optical image of the wearable ECG device based on the SEIC patch, worn on the subject's chest. (k) Representative ECG signals recorded by the wearable device during thirteen daily activities: (1) meditation, (2) breakfast, (3) brushing teeth, (4) walking, (5) working on a computer, (6) lunch, (7) working in a laboratory, (8) exercise, (9) cooking, (10) dinner, (11) showering, (12) watching TV, and (13) sleeping. (l) SNR of ECG signals recorded by the wearable device across the same 12 daily activities.

The self‐maintaining microenvironment at the skin‐electrode interface created by the SEIC patch and skin also contributed to improved ECG acquisition under motion artifacts. This effect was demonstrated by comparing the SEIC patch with Au‐coated fabric, 3 M, and Kendall electrodes. All four types of electrodes were simultaneously attached to the subject's forearms to capture ECG signals during various activities, including sitting, walking (1.5 steps s⁻¹), running (2 steps s⁻¹), stair‐climbing (2 steps s⁻¹), and cycling (2 cycles s⁻¹). The captured ECG signals were bandpass‐filtered at 0.5–3 Hz to extract the relevant components (Figure S11). These results show that ECG signals collected using the SEIC patch under motion artifacts exhibited less low‐frequency baseline drift than those collected by the Au‐coated fabric, 3 M, and Kendall electrodes. SNR analysis in Figure 3e confirms that the SEIC patch offers high dynamic noise resistance compared to the other electrodes. During walking, and running, stair‐climbing, the SEIC patch consistently maintained recognizable ECG waveforms, with an average SNR of approximately 7.5 dB (Figures S12–S14 and Figure 3e). In contrast, the Au‐coated fabric electrode failed to detect ECG signals, and the 3 M and Kendall electrodes were unable to preserve clear waveforms, with SNRs around ≈3.2 dB (Figures S12–S14 and Figure 3e). During cycling, only the Au‐coated fabric electrode showed degraded signal quality, while the SEIC patch, 3 M, and Kendall electrodes all produced ECG signals with waveforms and SNRs comparable to those recorded at rest (Figure S15 and Figure 3e). Moreover, we also did motion artifact ratio (MAR) analysis. This metric was computed for ECG signals across all electrode types (3 M, Kendall, SEIC, and Au‐coated fabric electrodes) under the full range of motion conditions tested in the study (walking, running, stair climbing, and cycling). The results (Figure S16) confirm that the SEIC patch consistently exhibited the lowest MAR across all signal types and motion conditions, quantitatively demonstrating its superior motion artifact rejection relative to the comparison electrodes.

While single‐lead ECG demonstrates basic signal quality, clinical cardiac monitoring typically requires multiple leads for comprehensive assessment. We therefore implemented a 3‐lead ECG configuration to validate the SEIC patch under clinically relevant conditions. SEIC patches were placed on the subject's chest and connected to the OpenBCI system to record Lead I and Lead II signals (Figure 3f). Performance was compared against 3 M and Kendall electrodes under conditions involving motion artifacts. The SEIC patch demonstrated robust signal acquisition in both Lead I and Lead II under motion artifacts (Figure 3g and Figure S17). As shown in Figure 3h, the SEIC patch consistently achieved high F1‐scores (above 0.9) across all activities, with minimal trial‐to‐trial variation. In contrast, 3 M and Kendall electrodes exhibited reduced performance, especially during dynamic activities such as walking, running, and stair climbing. During running, both commercial electrodes failed to detect ECG signals in both leads (Figures S18 and S19), and their F1‐scores dropped significantly under motion artifact conditions (Figure 3h).

To evaluate performance consistency across different users, we conducted a multi‐subject validation with ten participants using the SEIC patch in a 3‐lead configuration. ECG signals from Lead I and Lead II were recorded under both stationary and motion conditions. As shown in Figure S20, the SEIC patch consistently preserved clear ECG waveforms across all subjects. The distributions of F1‐scores and accuracies across participants and activities are shown in Figure 3i. The SEIC patch achieved F1‐scores and accuracies ranging from 0.85 to 1.0, with narrow distribution ranges, indicating reliable and reproducible performance across diverse users and movement conditions. This robustness demonstrates the effectiveness of the SEIC patch's self‐maintaining interface across diverse users.

To demonstrate real‐world implementation, we integrated the SEIC patch into a complete wearable ECG device for continuous daily monitoring. The device architecture shown in Figure S21 was worn on the chest and connected to an OpenBCI module mounted on the upper arm (Figure 3j). ECG signals were recorded throughout the day during thirteen typical daily activities: (1) meditation, (2) breakfast, (3) brushing teeth, (4) walking, (5) working on a computer, (6) lunch, (7) working in a laboratory, (8) exercise, (9) cooking, (10) dinner, (11) showering, (12) watching TV, and (13) sleeping. Representative ECG signals captured during these activities are shown in Figure 3k and Figure S22. The device successfully recorded clear ECG signals, consistently displaying distinguishable P, QRS, and T waveforms. To quantify signal quality, SNR values were calculated for all thirteen activities (Figure 3l). The SEIC‐based ECG device achieved strong SNR performance, with values ranging from approximately 8 to 18 dB and most activities exceeding 12 dB. These results confirm that the SEIC‐based system is capable of reliable, high ECG signal acquisition for continuous, real‐world health monitoring.

2.4. EOG and EMG Acquisition for Facial Expression Reconstruction in VR Headset

The self‐maintaining interface developed for motion‐robust cardiac monitoring is broadly applicable to biosensing scenarios where stable electrode—skin contact is challenging under dynamic conditions. To demonstrate this generalizability, we applied the SEIC patch to facial biosignal acquisition in VR headsets. VR systems, widely used in gaming, training, and remote collaboration, feature foam facial interfaces that contact the forehead, temples, and cheeks—regions suitable for recording EOG and EMG signals associated with eye movements and facial expressions. Reliable biosignal acquisition in this environment is challenging due to pressure variations from the foam interface, facial deformation during expressions, and heat and moisture accumulation during prolonged wear. By integrating SEIC patches into the VR facial interface, we show that the self‐maintaining microenvironment stabilizes electrode—skin contact and enables reliable acquisition of EOG and EMG signals during facial expressions. This provides sufficient signal fidelity for downstream applications such as real‐time avatar control, serving as a proof‐of‐concept demonstration of SEIC‐enabled wearable biosensing in VR, rather than a dedicated facial tracking framework.

For fundamental study, the SEIC patches were attached to the forehead with the following electrode configuration: the reference electrode (RE) was placed at the glabella, two working electrodes (WE) were positioned over the left and right frontalis, and the ground electrode (GE) was placed at the ear loop (Figure 4a) to capture of EOG and EMG signals during eye blinking and teeth griding. For comparison with Au‐coated fabric, 3 M, and Kendall electrodes, each of these electrodes was individually attached over the frontalis areas with SEIC patch to simultaneously measure EOG and EMG signals. The representative EOG and EMG signals obtained using the SEIC patch, Au‐coated fabric, 3 M, and Kendall electrodes are shown in Figure 4b,c, respectively.

FIGURE 4.

EOG and EMG acquisition under facial expressions. (a) Optical of a participant wearing SEIC patches positioned on the forehead for EOG/EMG signal acquisition. Representative EOG (b) and EMG (c) signals captured using four types of electrodes (SEIC patch, Au‐coated fabric, 3 M, and Kendall) under eye blinking, and teeth grinding, respectively. SNR comparison of EOG (d) and EMG (e) signals for various body movements (sit, step in place, arm raise‐lowering, arm cross, head shake, head nod) recorded using SEIC patch, Au‐coated fabric, 3 M, and Kendall electrodes. (f) Image of the SEIC patch integration on the facial interface of the Meta Quest 3 headset, including six working electrodes (WE1–WE6), and one reference electrode (RE). (g) Facial interface of Headset with OpenBCI board mounted for signal acquisition. (h) Optical image of a participant wearing the integrated system with SEIC patches and data acquisition board. (i) SNR comparison of EMG signals during nine facial expressions (fear, happiness, anger, sadness, surprise, contempt, disgust, pain, amusement) across six working electrodes (WE1–WE6), demonstrating spatial signal response on the face.

Having confirmed basic signal acquisition, we next evaluated whether the SEIC patch could maintain this performance during the dynamic conditions typical of VR use. VR users rarely remain still—they turn their heads to explore virtual environments, gesture to interact with objects, and increasingly walk in place for locomotion. To simulate these real‐world conditions, we recorded EOG and EMG signals while participants performed representative movements: sitting, stepping in place (2 steps s⁻¹), arm raising and lowering (1 cycle s⁻¹), arm crossing (1 cycle s⁻¹), head shaking (1 cycle s⁻¹), and head nodding (1 cycle s⁻¹). These results were compared with those obtained from Au‐coated fabric, 3 M, and Kendall electrodes. The signal‐to‐noise ratio (SNR) analysis results shown in Figure 4d,e indicate that the SEIC patch exhibits a higher SNR than the Au‐coated fabric, 3 M, and Kendall electrodes. Although the SNR of all electrodes decreased under motion artifacts, the SEIC patch maintained a relatively high SNR, whereas the SNR of the Au‐coated fabric, 3 M, and Kendall electrodes dropped significantly. These results further confirm the role of the high‐humidity microenvironment created between the SEIC patch and the skin in enhancing biosignal acquisition under motion artifacts.

To validate real‐world performance of SEIC patches in a VR headset, we integrated the patches into the facial interface of the Meta Quest 3 headset. The configuration included one RE, six WEs (Figure 4f), and an additional common GE placed on the ear loop. These electrodes were connected to an OpenBCI board mounted on the top of the facial interface (Figure 4g). This integrated facial interface was conformally worn by the user to enable real‐time biosignal acquisition during various facial expressions (Figure 4h). The biosignals recorded by the six WEs during nine facial expressions (fear, happiness, anger, sadness, surprise, contempt, disgust, pain, and amusement) are shown in Figure S23. The SNR analysis for each electrode across different expressions is presented in Figure 4i. The results demonstrate high and diverse signal quality from all six electrodes, ensuring that different facial muscle groups and eye movements are effectively captured.

The distinct biosignal patterns and varying SNR values observed across different facial expressions revealed an intriguing possibility that these signals contain sufficient spatiotemporal information to go beyond simple expression classification. The richness and consistency of the signal variations across electrode positions suggested that the biosignals could potentially encode the continuous deformation of facial geometry itself. This observation motivated us to explore whether we could directly reconstruct 3D facial expressions from these biosignal patterns, essentially creating a mapping from the electrical activity of facial muscles to their corresponding geometric configurations.

Such biosignal‐driven facial reconstruction enables real‐time avatar animation in VR without the need for cameras, providing an alternative pathway for facial tracking. In our approach, biosignal‐based reconstruction captures facial muscle activity through surface electrodes integrated directly into the VR facial interface. This method requires no optical components or additional processing hardware and maintains robust performance across headset movement and user variability. To demonstrate this capability enabled by the SEIC interface, we developed a cross‐modal learning framework that maps multichannel biosignals to 3D facial landmarks using synchronized biosignal–video data from nine facial expressions (Figure 5a). A 1D convolutional neural network was trained to learn the relationship between temporal biosignal patterns and facial geometry represented by 86 landmarks covering key facial regions, and was evaluated on unseen test data to ensure generalization (Figure 5b). The system achieved high reconstruction accuracy across all facial regions (Figure 5c), with mean absolute errors ranging from 0.49 mm (right eye) to 1.97 mm (lower lip), and sub‐millimeter accuracy in most eye and eyebrow regions. These results are comparable to state‐of‐the‐art camera‐based facial tracking systems (typically 1–2 mm error), while relying solely on biosignals without any visual input. The reconstructed 3D facial animations further demonstrate the ability to capture both global expression changes and fine‐grained muscle dynamics across diverse emotions, including neutral, happiness, sadness, contempt, and amusement (Figure 5d). Continuous reconstruction shows smooth temporal transitions and stable performance throughout expression sequences (Video S1). The multichannel biosignal patterns provide sufficient spatial information to distinguish similar expressions involving overlapping facial muscle groups. Importantly, this reconstruction is presented as a proof‐of‐concept demonstration enabled by the SEIC interface, rather than a standalone facial tracking framework.

FIGURE 5.

Reconstructed 3D facial animation based on biosignals. Schematic diagrams illustrate the processing steps in the cross‐modal supervised learning framework, including: (a) the training phase, and (b) the testing phase. (c) The average MAE for key facial regions. (d) The reconstructed 3D animation of five facial expressions compared with ground truth.

3. Conclusions

In this work, we demonstrate that the long‐standing challenge of achieving a stable electrode–skin interface can be addressed through an innovative strategy—engineering a self‐maintaining humid microenvironment that actively sustains the conditions required for ionic conduction. Rather than relying on static materials to preserve contact, our SEIC patch establishes a dynamic interface in which evaporative cooling continuously regenerates the ionic pathway between the skin and the electrode. Despite decades of research into electrode materials and coatings, we believe this concept of integrating humidity and temperature regulation directly into electrode design represents a fundamental breakthrough with strong potential for translation and commercialization.

The key insight underlying this approach is that a humid microenvironment can form a self‐sustaining electrochemical interface. The Au‐coated fabric integrated with the hydrogel cooling patch absorbs and retains moisture, producing a hydrated electrode surface that facilitates ion transport and enhances skin coupling. When attached to human skin, the SEIC patch induces evaporative cooling, lowering the local skin temperature and creating a high‐humidity microenvironment. The resulting thin condensate layer bridges the electronic‐to‐ionic conduction gap, forming a continually renewed ionic conduction pathway. The observed 200‐fold reduction in impedance compared with dry Au‐coated fabric electrodes confirms that this condensation layer effectively eliminates the high‐resistance air gaps that typically limit bioelectronic interfaces. Unlike conventional gel electrodes, where the ionic medium degrades through dehydration, our interface self‐regenerates through continuous vapor–liquid phase transition.

This self‐maintaining mechanism marks a departure from traditional approaches to bioelectronic interface design. Previous strategies have primarily focused on developing more conductive polymers, more stable gels, or more compliant substrates—‐materials‐based solutions aimed at preserving a static interface under dynamic conditions. In contrast, our work shows that the interface itself can be transformed into an active system that autonomously maintains its electrochemical environment. The exceptional stability observed across 2000 attachment–detachment cycles, eight days of continuous ECG monitoring, and consistent performance across diverse users and motion conditions does not stem from unusual material robustness, but rather from the interface's ability to continually re‐establish its ionic conduction layer. Each time the electrode recontacts the skin, the cooling process restarts condensation, effectively regenerating a fresh electrochemical junction.

The implications extend beyond enhanced electrical performance. By maintaining a stable ionic environment at the skin interface, the SEIC patch enables reliable biosignal acquisition under conditions that typically introduce severe motion artifacts. The preserved ECG, EOG, and EMG signal quality during vigorous movement suggests that the condensed water layer provides mechanical decoupling between skin deformation and electrode displacement. This layer can dynamically redistribute during motion, maintaining ionic continuity even as the mechanical interface stretches or shifts – a property that rigid gels or dry electrodes inherently lack.

The SEIC patch's ability to establish and maintain a self‐renewing bioelectronic interface has broad applications. In clinical contexts, SEIC patches could enable continuous ECG monitoring during cardiac rehabilitation exercises—precisely when arrhythmia risk is highest, but traditional electrodes often fail due to motion artifacts. For neuromuscular disorder assessment or stroke rehabilitation, motion‐robust signal acquisition could facilitate quantitative tracking of motor recovery during active therapy sessions. Beyond medical applications, the principle of microenvironment engineering at the skin–electrode interface may transform long‐term biosensing for athletic performance monitoring and human–computer interaction. The integration of SEIC technology with VR systems for 3D facial reconstruction is particularly promising: as headsets become lighter and more prevalent in training, telepresence, and social communication, the ability to capture authentic facial expressions without cameras will be crucial for natural interaction and emotional realism.

More broadly, this work demonstrates that engineering the local environment where electronics meet biology can provide a more robust and generalizable pathway to overcome interfacial instability than materials innovation alone. This paradigm – of dynamic, self‐regulating interfaces—may redefine how future bioelectronic systems achieve long‐term reliability and comfort in continuous human use.

4. Experimental Section

To establish a continuous and multimodal biosensing platform, we detail the fabrication of the SEIC patch, the development and evaluation of a continuous ECG monitoring system, and the extension of SEIC electrodes to real‐time EOG and EMG monitoring for facial expression analysis.

4.1. Materials

The Au‐coated fabric consists of gold coated on a base material of silver‐coated nylon. The commercial cooling patch including: Gelling agent: Hydrophilic macromolecular material (sodium polyacrylate). Cooling agents: Mint and borneol. Moisturizing components: Glycerin. And, water: The primary component of the hydrogel responsible for absorbing and dispersing heat through evaporation, providing the physical cooling effect.

4.2. Fabrication of SEIC Patch

The Au‐coated fabric electrode was cut into a circular shape with a diameter of 8 mm, similar to the size of a Ag/AgCl commercial electrode, from an Au‐coated fabric sheet. It was then cleaned sequentially in acetone, ethanol, and deionized (DI) water (resistivity: 18.0 MΩ cm, purchased from ACME Analytical Solutions Inc.) using sonication for 5 min in each solution. After cleaning, the electrode was dried on a hot plate at 120°C for 30 min. This temperature is well below the melting point of Au and below the thermal processing temperatures typically tolerated by nylon‐based fabric substrates. The Au‐coated fabric is a commercially supplied material. Information regarding Au film thickness and any proprietary interfacial or adhesive layers is not available from the manufacturer. No delamination, cracking, or degradation of the metal coating, nor any degradation of the electrical conductivity of the Au‐coated fabric, was observed after drying at 120°C for 30 min. A copper (Cu) connection wire was then soldered to the center of the Au‐coated fabric electrode. For integration with the hydrogel cooling patch, the hydrogel sheet was cut into 20 × 20 mm squares, and a 0.5 mm hole was punched at the center of each patch. The connection wire was passed through the hole to integrate the Au‐coated fabric electrode with the hydrogel cooling patch. Hot glue was applied to secure the connection wire to the hydrogel patch. A foam‐based moisture‐resistant layer was then applied around the hydrogel patch to reduce vapor evaporation. Finally, a waterproof tape was placed over the top of the cooling patch and foam layer to hold them together. The SEIC electrode is based on commercially available cooling patch and Au‐coated fabric. Therefore, the proposed structure is inherently compatible with scalable production methods similar to those used for existing commercial cooling patches and Au‐coated fabric. However, we acknowledge that the current study is performed at a laboratory scale, and full industrial‐scale optimization, including long‐term shelf stability, packaging, and automated manufacturing yield, remains to be systematically investigated in future work.

4.3. Commercial Electrodes

The 3 M Red Dot electrode (2560) has an overall size of 40 × 35 mm, a Ag/AgCl sensor with a diameter of 8 mm, and an oval‐shaped gel area of 19 × 17 mm. The Kendall electrode (Cardinal Health) has an overall size of approximately 30 × 24 mm, an Ag/AgCl sensor diameter of 8 mm, and a gel diameter of 16 mm. These electrodes were selected to compare with SEIC because they are widely used, commercially available, and represent standard benchmarks for skin‐electrode impedance and signal quality. This allows for a meaningful comparison of the SEIC patch against clinically validated electrodes.

4.4. Temperature and Moisture Measurement

The temperature and moisture content of SEIC electrodes and standalone Au‐coated fabric electrode were measured by using laser thermometer and humidity probe (Sper Scientific 800021C), respectively. Each measurement condition was replicated three times. In case of skin temperature and skin moisture, the sin on front arm was measured by also using laser thermometer and humidity probe before and after attachment SEIC electrodes and standalone Au‐coated fabric electrode. Each measurement condition was replicated three times.

4.5. FE‐SEM Measurement

The Au‐coated fabric was cut into 1 by 1 cm. It was then cleaned sequentially in acetone, ethanol, and DI water using sonication for 5 min in each solution. After cleaning, the electrode dried on a hot plate at 120°C for 30 min. The drying step was performed to remove residual moisture prior to FE‐SEM characterization. After dried Au‐coated fabric is being loaded into the chamber for analysis by FE‐SEM (JEOL, Model JSM‐6500F). Images were acquired at an accelerating voltage of 15 kV.

4.6. Impedance Measurement

To compare the skin contact impedance of SEIC electrodes with standalone Au‐coated fabric, 3 M, and Kendall electrodes, all electrodes were attached to the forearm. Each measurement condition was replicated three times. Impedance was measured using a custom‐built system over a frequency range of 200 Hz to 100 kHz. A schematic and detailed description of the custom‐built system are provided in Figure S24.

4.7. Continuous ECG Monitoring System and Evaluation Framework

The continuous ECG monitoring system is built around an OpenBCI Cyton board. This device provides 24‐bit resolution at a 250 Hz sampling rate, ensuring high‐fidelity physiological recordings. Acquired signals are transmitted wirelessly to a nearby host computer via Bluetooth, where they undergo cleaning and pre‐processing to remove noise and extract physiologically meaningful ECG, EOG, and EMG features. These processed signals are then directed into application‐specific pipelines for continuous ECG monitoring and facial animation. To validate the robustness and utility of the SEIC patch within this system, we employed a progressive evaluation strategy, beginning with controlled laboratory conditions, extending to full‐day deployments in naturalistic environments, and culminating in multi‐day continuous measurements. Each measurement condition was replicated three times.

Step 1—Controlled laboratory validation: long‐term baseline study: We first assessed the stability of ECG acquisition under conditions where motion artifacts were minimized. A single participant (male, 44 years) wore three electrode types in parallel, including SEIC patch, 3 M, and Kendall, placed in a row on both forearms. ECG signals from all electrodes were recorded simultaneously to allow direct comparison. Recordings were performed for one minute each morning at 9:00 a.m. over a period of eight consecutive days. During this period, the electrodes remained continuously attached to the skin for 24 h per day, encompassing a wide range of normal activities including eating, working, walking, driving, showering, and sleeping. This design enabled a head‐to‐head evaluation of SEIC patches against commercial electrodes across multiple days of continuous use, providing insight into both stability and wearability.

Step 2—Evaluation under motion artifacts: activity‐based protocol: To examine the resilience of SEIC patches to movement‐induced artifacts, we conducted following studies: (i) Single‐lead ECG: A participant wore four electrode types simultaneously, including SEIC patch, Au‐coated fabric, 3 M, and Kendall, arranged in parallel on both forearms and recorded ECG signal under a range of physical activities that introduce progressively more challenging motion conditions: sitting at rest, walking at 1.5 steps s⁻¹, running at 2 steps s⁻¹, stair climbing at 2 steps s⁻¹, and cycling at 2 cycles s⁻¹. (ii) Three‐lead ECG: A participant worn SEIC patches, 3 M, and Kendall electrode in a standard 3‐lead ECG configuration on the chest, sequentially, and recorded ECG signal under the same activities. (iii) For 3‐lead ECG (multiple participants): Ten healthy participants (ages 26–44). Each participant wore SEIC patches configured in a standard 3‐lead setup, and ECG signal were recorded throughout the same activity set. This experimental design enabled a direct evaluation of the SEIC patch against multiple commercial and fabric‐based electrodes, highlighting differences in robustness during motion‐intensive scenarios.

Step 3—Daily naturalistic monitoring: full‐day protocol: To evaluate the SEIC patch in everyday contexts, we implemented a comprehensive daily protocol spanning both routine and physically demanding activities. A participant was instructed to continuously wear an SEIC patch‐based ECG monitoring device for the entirety of a day, allowing uninterrupted signal capture across diverse activity states. ECG signals were recorded during thirteen representative activities distributed across the daily timeline:

• Morning: meditation, breakfast, brushing teeth, walking to the bus, and working on a computer.

• Noon/Afternoon: lunch, laboratory work, and structured physical exercise.

• Evening/Night: cooking, dinner, showering, watching television, and sleeping.

This was designed to replicate real‐world conditions in which electrode reliability is frequently challenged, such as during personal hygiene or exercise to offer a stringent test of signal quality under naturalistic circumstances.

4.8. Real‐Time Continuous EOG and EMG Monitoring

In addition to ECG monitoring, SEIC patches were evaluated for their capacity to capture EOG and EMG signals, thereby extending their functionality to real‐time monitoring of eye and facial muscle activity. These signals provide an important foundation for applications such as fatigue detection, teeth‐grinding assessment, and facial expression recognition. Each measurement condition was replicated three times.

Step 1— Baseline signal validation under motion: To directly compare SEIC patches against alternative electrode technologies, we conducted paired experiments with Au‐coated fabric, 3 M, and Kendall electrodes. For each pairing (SEIC/Au‐coated fabric, SEIC/3 M, SEIC/Kendall), electrodes were tested in two reciprocal configurations: (i) SEIC on the left frontalis with the comparison electrode on the right frontalis, and (ii) reversed positions. During each configuration, a participant performed a set of controlled activities designed to introduce motion: sitting at rest, stepping in place (2 steps s⁻¹), arm raising/lowering (1 cycle s⁻¹), arm crossing (1 cycle s⁻¹), head shaking (1 cycle s⁻¹), and head nodding (1 cycle s⁻¹). EOG and EMG signals were simultaneously recorded during eye‐blinking and teeth‐grinding events to evaluate motion tolerance and stability of acquisition across electrode types.

Step 2—Facial expression monitoring protocol: Building on robustness observed in baseline tests, we next evaluated SEIC‐based EOG and EMG signals for their ability to support real‐time facial expression monitoring. A participant wore a facial interface embedded with SEIC patches and sat in front of a laptop with an integrated camera, which served both as a ground‐truth recording device and as a training data source. The participant was instructed to repeatedly perform nine distinct facial expressions 50 times each, holding every expression for 3 s and returning to a neutral facial state for 5 s between repetitions. To guide consistency, the corresponding target expression image was displayed on the laptop screen during data collection.

Step 3—Multimodal data preprocessing: We design the preprocessing pipeline to align physiological recordings with video‐based facial landmarks:

• ○Video Segmentation and Landmark Extraction: Input videos were truncated to a maximum of 8 s, sampled at 30 frames per second, and segmented into 60‐frame overlapping clips with a stride of 45 frames (25% overlap). Facial landmarks were extracted using MediaPipe Face Mesh, which outputs 468 keypoints per frame. To reduce dimensionality while preserving emotional salience, we selected 86 landmarks corresponding to the eyebrows, eyes, nose, lips, and chin.
• ○Signal Segmentation and Filtering: Physiological recordings were segmented into overlapping windows of 500 samples with a stride of 375 samples (25% overlap). Each segment was passed through a bandpass filter to remove low‐frequency drift and high‐frequency noise, retaining the components most relevant for affective state recognition.
• ○Multimodal Alignment and Labeling: Each (video, signal) segment pair was temporally aligned and labeled with the corresponding target expression. File indexing ensured consistent cross‐modal matching (e.g., *1.mp4 with *1.csv).
• ○Data Splitting: The dataset was pooled across all emotion classes and partitioned into training (≈75%), validation (≈12.5%), and test (≈12.5%) sets using stratified sampling to preserve class balance.

Step 4—Deep learning model for landmark regression: We propose a multimodal deep learning framework to regress 3D facial landmarks directly from biosignal time‐series data. The model is implemented as a deep 1D CNN that captures spatio‐temporal dependencies in EOG and EMG signals.

• ○Network Architecture: The proposed model processes input sequences of shape (C, T), where C is the number of input channels corresponding to the biosignal modalities, and T is the temporal sequence length. The architecture consists of four successive convolutional blocks, each followed by batch normalization, leaky ReLU activation (negative slope = 0.3), and max pooling. Dropout is applied after each of the first three blocks to improve generalization. Let F be the number of base convolutional filters and k be the filter width. The network blocks are as follows:

  1. Conv Block 1: Conv1D(C, F, k) → BatchNorm1D(F) → LeakyReLU → MaxPool1D

  2. Conv Block 2: Conv1D(F, 2F, k) → BatchNorm1D(2F) → LeakyReLU → MaxPool1D

  3. Conv Block 3: Conv1D(2F, 4F, k) → BatchNorm1D(4F) → LeakyReLU → MaxPool1D

  4. Conv Block 4: Conv1D(4F, 8F, k) → BatchNorm1D(8F) → LeakyReLU

After the final convolutional layer, the feature maps are flattened and passed through two fully connected layers:

• FC1: Linear layer mapping to 400 units, followed by LeakyReLU and dropout.

• FC2: Linear layer projecting to the final output dimension of 3L, where L is the number of 3D facial landmarks.

• ○
  Loss Function: The Wing Loss function is employed as the regression objective, as it improves robustness to outliers and increases accuracy for small errors, which is a critical requirement for precise landmark localization. The WingLoss is designed as follows:

  $$\mathrm{WingLoss}(y,\widehat{y)}=\frac{1}{N}\sum _{i=1}^N{l}_i$$

  $$l_i=\left\{\begin{array}{c}\omega \ast \log \left(1+\frac{\left|y_i-\widehat{y_i}\right|}{\epsilon }\right),if\left|y_i-\widehat{y_i}\right|\\ \left|y_i-\widehat{y_y}\right|-Cotherwise\end{array}\right.$$

  $$C=\omega -\omega \ast log\left(1+\frac{\omega }{\epsilon }\right)$$

where y_i is the ground‐truth landmark coordinate, $\widehat{y_i}$ is the predicted landmark coordinate, and N is the number of landmarks, ω (omega) is a hyperparameter controlling the range of nonlinearity, and ϵ (epsilon) is a hyperparameter controlling the curvature of the nonlinearity. Standard L2 loss tends to overemphasize large errors, while L1 loss is not informative enough, making it harder to refine predictions once coarse alignment is achieved.

In designing our deep learning framework, we explored various architectures that we hypothesized would improve the temporal abilities of the model. In particular, we attempted to add gated recurrent unit (GRU) layers and long short‐term memory (LTSM) layers, which are widely known to capture sequential dependencies, mitigate vanishing gradient issues, and efficiently model long‐term temporal patterns. However, we found that a simple architecture with just 1D CNN layers worked the best as it prevented the model from overfitting and generalized well. In the end, we observed that a simpler architecture composed of 1D convolutional layers outperformed these recurrent variants. This convolutional approach reduced the risk of overfitting and yielded stronger generalization across emotions and expressions. We also explored augmenting the input biosignal data with emotion labels encoded as embeddings. However, this led the model to overemphasize the label embeddings at the expense of the biosignal features, ultimately reducing predictive performance.

Author Contributions

T. Q. T., P.N., and D.G. conceived the idea and designed the study. T.Q.T. designed and characterized the SEIC patch. T.Q.T. and A.R. established the connection between the SEIC patch and the OpenBCI module. T.Q.T. characterized the SEIC patch's capability in biopotential signal acquisition under both stationary and motion conditions. T.Q.T. also integrated the SEIC patch into the facial interface of the VR headset and evaluated biosignal acquisition during various facial expressions. Z.L. designed and fabricated the electronic circuit for measuring electrode–skin contact impedance. S.K.S. proposed and implemented the algorithm for biosignal processing and 3D facial reconstruction. T.Q.T., S.K.S., V.P., and D.G. discussed the results and wrote the paper. All authors reviewed and provided feedback on the final manuscript.

Ethics Statement

Subjects between the ages of 26 and 44 were eligible to participate in this study. The study was approved by the Institutional Review Board of the University of Massachusetts Amherst (IRB: 5481) to ensure the protection of human participants. The approved protocol covers all aspects of the human studies described in the Experimental Section. Prior to participation, all subjects provided written informed consent. During the consent process, a member of the study team reviewed the entire consent form and provided a detailed explanation of the study's purpose, potential risks and discomforts, data sharing, privacy safeguards, and the possibility that photos or videos of participants might be taken during the study. Study staff then explained the specific activities that participants would be asked to complete. Participants who agreed to take part were provided with a copy of the signed consent form. In total, n = 10 males were recruited for the field study.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Supporting File: adhm71382‐sup‐0001‐SuppMat.docx.

Supporting Video: adhm71382‐sup‐0002‐VideoS1.mp4.

Data Availability Statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.