Nanoporous Sleep Tracking Patch

Determine how the anatomy of the retroauricular–cervical region constrains signal quality, and identify the electrode layout that maximizes it. Volume-conductor modeling and a small-scale measurement study with conventional electrodes will compare candidate layouts against reference EEG, ECG, and EDA. Electrode spacing and orientation, reference and ground placement, muscle interference, hair, and skin properties will be varied or measured.

Working hypothesis: at least one layout will meet pre-specified quality criteria for EEG and ECG simultaneously.

Identify materials that allow a breathable, skin-safe, electrically stable patch. Candidate porous substrates, electrode materials, and skin adhesives will be screened on the bench for breathability, flexibility, chemical stability, and biocompatibility, then tested on skin. Shortlisted materials will be combined into small-format test patches recorded alongside conventional electrodes.

Working hypothesis: a porous substrate with a suitable electrode–adhesive combination will maintain stable electrode–skin contact over 8 h without skin irritation.

Build a pipeline that quantifies the measurements and estimates the sleep information available at this site. Signal-quality scoring, artifact separation, and feature extraction will be developed for the data from Aims 1 and 2. Using public PSG datasets, models restricted to the signals a patch could record will estimate how well sleep stages, including REM sleep, can be identified.

Working hypothesis: combining EEG with cardiac and muscle-tone features will improve REM sleep identification over EEG alone.

The expected outcome is an evidence-based design specification for a first prototype, including electrode layout, materials, and analysis pipeline, together with a documented account of the anatomical and technical limits of single-site recording. These results will set up subsequent prototype development and overnight validation against PSG, aimed at REM sleep tracking and more comfortable clinical sleep studies.

Under the American Academy of Sleep Medicine (AASM) rules, sleep is scored in 30-s epochs as wake (W), NREM stages N1, N2, and N3, or REM (R), using EEG, electrooculography (EOG), and chin electromyography (EMG) (Berry et al., 2020). Each stage has characteristic cortical signatures: posterior alpha activity (8–12 Hz) during relaxed wakefulness, sleep spindles (11–16 Hz) and K-complexes in N2, high-amplitude slow waves (0.5–2 Hz) in N3, and low-amplitude mixed-frequency activity accompanied by rapid eye movements and muscle atonia in REM sleep. The autonomic nervous system changes in parallel. Sympathetic outflow declines and parasympathetic (vagal) influence on heart rate increases with deepening NREM sleep, whereas REM sleep is marked by sympathetic surges and greater cardiovascular variability (Somers et al., 1993; Trinder et al., 2001). Electrodermal activity follows a distinct pattern: bursts of skin conductance responses occur most frequently during slow-wave sleep and are comparatively rare during REM sleep (Boucsein, 2012; Sano et al., 2014). Because cortical and autonomic signatures carry partially independent information, combining them may resolve stages that are ambiguous from any single signal. REM sleep is of particular interest to theories linking sleep physiology to dreaming and memory (Hobson et al., 2000), yet its identification depends on features, such as eye movements and muscle atonia, that are difficult to measure outside the laboratory.

PSG remains the reference standard, but it requires trained technicians, many wired sensors attached to the scalp, face, chest, and limbs, and an unfamiliar environment that can itself alter sleep. For patients, particularly children, older adults, and people with anxiety or sensory sensitivities, this setup can be uncomfortable and can degrade the very sleep being measured. It is also rarely repeated across nights. Consumer devices offer the opposite trade-off. Validation studies indicate that wrist-worn and other consumer sleep trackers generally classify sleep versus wake well but are less accurate in estimating wake after sleep onset and in discriminating individual sleep stages (de Zambotti et al., 2019; Chinoy et al., 2021). Because these devices do not record cortical activity, their staging relies on indirect correlates such as movement and heart rate.

Several groups have shown that EEG can be recorded from in and around the ear (Looney et al., 2012; Debener et al., 2015), and behind-the-ear electrode arrays have supported automatic sleep staging with moderate-to-substantial agreement with PSG-based scoring (Mikkelsen et al., 2017; Sterr et al., 2018). The region spanning the upper neck and the area behind the ear is anatomically complex. It overlies the mastoid process and the temporal and occipital bones, which attenuate and spatially blur cortical signals; it lies close to the sternocleidomastoid, trapezius, and posterior auricular muscles, whose activity produces EMG with a broad spectrum that overlaps substantially with EEG and is most prominent at temporal sites (Goncharova et al., 2003); it varies in hair coverage, skin thickness, and sebaceous activity across individuals; and it is subject to pressure and shear when a person lies on a pillow. Each of these factors affects signal quality and electrode contact. Some may also be informative: loss of muscle tone is a defining feature of REM sleep, so sensitivity to cervical EMG may carry information useful for REM detection if cortical and muscular sources can be separated.

Heart rate variability (HRV) provides a noninvasive index of autonomic regulation. Time-domain indices such as the root mean square of successive differences (RMSSD) and frequency-domain indices such as high-frequency (HF; 0.15–0.40 Hz) power primarily reflect vagal modulation, whereas low-frequency (LF; 0.04–0.15 Hz) power reflects a mixture of sympathetic and parasympathetic influences (Task Force of the ESC and NASPE, 1996; Shaffer and Ginsberg, 2017). Accurate HRV analysis requires precise beat-to-beat (R–R) intervals, best obtained from ECG. EDA reflects eccrine sweat gland activity under sympathetic cholinergic control and therefore complements HRV as a sympathetic measure (Boucsein, 2012). Most EDA research uses the palms or fingers, where eccrine gland density is high; the neck and retroauricular regions have lower gland density, so EDA feasibility at this site is treated as an open question. All three measurements also depend on the biochemistry of the electrode–skin interface. The stratum corneum, its hydration, and the electrolyte content of sweat determine electrode–skin impedance and electrode potentials, and EDA in particular requires nonpolarizing electrodes and a stable electrolyte interface (Boucsein, 2012).

Epidermal electronics established that ultrathin, stretchable devices can conform to the skin and record electrophysiological signals with mechanical properties similar to those of skin (Kim et al., 2011). Subsequent work demonstrated gas-permeable nanomesh conductors that can be worn for extended periods with minimal inflammatory response, highlighting the importance of breathability for long-term wear (Miyamoto et al., 2017). Occlusive materials trap sweat and moisture, which can degrade adhesion, alter impedance, and irritate skin over a full night. For a sleep patch, breathability and biocompatibility are therefore requirements for both safety and signal stability. More broadly, recent clinical work on AI-integrated wearable sensors, such as wearable ultrasound for noninvasive estimation of central venous pressure, illustrates the path from sensor design to clinical validation that this project aims to follow (Dong et al., 2026).

Deep learning models trained on one or a few EEG channels can approximate expert sleep scoring (Supratak et al., 2017; Perslev et al., 2021). Two issues are critical for wearable data. First, wearable recordings contain artifacts from movement, muscle activity, and changing electrode contact that are uncommon in curated PSG datasets. Second, models frequently perform worse on individuals outside the training set, so evaluation must be strictly subject-independent. Large public PSG datasets (Kemp et al., 2000; Goldberger et al., 2000; Zhang et al., 2018) make it possible to estimate, before a device exists, how much sleep information is retained when only a restricted set of signals is available.

This R&D phase addresses the questions that must be answered before a multimodal sleep patch can be designed responsibly. The results have value regardless of outcome: a positive result provides a design specification grounded in anatomy, materials testing, and quantitative analysis, whereas a negative or partial result defines the physical limits of single-site recording and indicates which signals require separate placement. In the longer term, a validated patch could support REM sleep tracking across many nights and make clinical sleep studies more comfortable and accessible by reducing the number of wired sensors patients must wear.

The three aims run in parallel and inform one another (Table 2). Aim 1 defines where and how electrodes should be placed. Aim 2 identifies the materials from which a patch can be made and produces small-format test patches. Aim 3 provides the analysis pipeline for the data from Aims 1 and 2 and estimates, using public datasets, how much sleep information the selected signals can carry. The first manuscript will combine a literature review of candidate anatomical sites with the Aim 1 modeling results. This phase does not build or validate a complete overnight device; that work is described under Future Directions.

Healthy adults aged 18–45 years will be recruited. Exclusion criteria are a diagnosed sleep disorder, cardiac arrhythmia, or neurological disorder; use of medications known to affect sleep or autonomic function; known skin disease at the test sites; and known sensitivity to medical adhesives. Habitual sleep quality will be assessed with the Pittsburgh Sleep Quality Index (Buysse et al., 1989). Recruitment will aim for balance across sexes and will include a range of skin types and hair characteristics, because these affect electrode contact at the retroauricular site. Planned sample sizes are 20 participants for the Aim 1 measurement study, 10 for Aim 2 skin-wear testing, and 8 for the Aim 2 test-patch pilot. Because this is an R&D phase, sample sizes are chosen to inform estimation and design decisions rather than to confirm hypotheses.

A structured literature review will summarize the anatomy of the region (bone, muscle, hair coverage, skin thickness, and sweat and sebaceous gland distribution) and published recordings of ear-EEG, neck-based ECG, and nonpalmar EDA, extracting signal amplitudes, signal-to-noise ratios, and artifact profiles by site. A realistic standardized head model (Huang et al., 2016) will be used to compute lead fields for candidate bipolar electrode pairs within the region and to estimate their sensitivity to cortical sources relevant to sleep, including frontocentral generators of slow waves and spindles and posterior generators of alpha activity. The output is a ranked shortlist of three to five candidate layouts.

• Design. Each participant will wear a dense grid of conventional Ag/AgCl electrodes over the candidate region on one side, allowing the shortlisted layouts to be reconstructed by re-referencing. Reference signals will be recorded concurrently: scalp EEG at standard 10–20 positions (Klem et al., 1999) referenced to the contralateral mastoid, chest-lead ECG, and palmar EDA measured exosomatically with a constant-voltage method. At each grid location, skin properties relevant to the electrode interface will be documented, including hair coverage, skin hydration, and baseline electrode–skin impedance.

• Protocol. Each session will include (a) resting recordings with eyes open and eyes closed to assess alpha reactivity; (b) controlled artifact tasks, including jaw clenching, swallowing, head rotation, and lying on each side with pillow pressure on the site; (c) sympathetic challenges known to elicit skin conductance responses, including deep inspiration, brief auditory tones, and mental arithmetic; (d) paced breathing to assess respiratory modulation of heart rate; and (e) a 90-min afternoon nap opportunity to capture sleep onset, spindles, and, where present, slow waves.

• Analysis. Signal processing for this aim is carried out with the pipeline developed in Aim 3. R-peaks will be detected with an established QRS detection algorithm (Pan and Tompkins, 1985), and phasic and tonic EDA components will be separated using continuous decomposition analysis (Benedek and Kaernbach, 2010). Agreement with reference measures will be summarized with Bland–Altman analysis (Bland and Altman, 1986) and intraclass correlation coefficients. Effects of layout and skin properties will be estimated with linear mixed-effects models with participant as a random intercept, with false discovery rate correction across outcomes.

• Quality criteria. Provisional criteria are listed in Table 1 and will be finalized after Aim 1a and preregistered before data collection. A layout is carried forward if it meets the EEG and ECG criteria; EDA performance is a secondary criterion.

Based on prior behind-the-ear recordings, EEG of usable quality is expected for at least one layout. ECG amplitude at the neck is expected to be lower than at the chest; if R-peak criteria are not met, electrode spacing will be extended inferiorly along the neck to better align with the cardiac axis. EDA is the signal at greatest risk because of lower sweat gland density. If EDA criteria are not met, two alternatives will be evaluated: a small satellite EDA electrode on a higher-density site connected to the main patch, or a design that omits EDA. If muscle contamination limits EEG quality, analyses will emphasize frequencies below 16 Hz and use EMG-referenced artifact regression.

• Substrates. Candidate porous substrates, including electrospun polymer nanofiber membranes, porous silicone elastomers, and microperforated medical-grade polyurethane films, will be characterized for thickness, stretchability, elastic modulus, and water vapor transmission rate measured according to ASTM E96.

• Electrodes. Thin-film Ag/AgCl, gold, and conductive-polymer electrodes will be compared, with and without a thin hydrogel interface. Because exosomatic EDA requires nonpolarizing electrodes (Boucsein, 2012), Ag/AgCl will be prioritized for EDA contacts. Electrodes will be tested for impedance spectra (1 Hz–1 kHz), electrode noise, and DC potential drift, and for stability after soaking in artificial sweat.

• Adhesives. Skin-contact adhesives will be evaluated for peel strength under dry and sweat-exposed conditions and for chemical compatibility with the electrode materials, including corrosion or delamination after prolonged contact.

• Biocompatibility. Only materials with existing skin-contact biocompatibility data will be advanced to human testing, and candidate assemblies will be evaluated against the relevant parts of the ISO 10993 series for skin-contacting devices (cytotoxicity, irritation, and sensitization) before wear testing.

Ten participants will wear small samples of shortlisted materials, without electronics, on the retroauricular–cervical site for up to 8 h. Outcomes will include standardized visual scoring of erythema and edema before and after wear, transepidermal water loss, a comfort rating on a visual analog scale, and the proportion of sample area remaining adherent at removal.

Shortlisted substrates, electrodes, and adhesives will be combined into small-format test patches implementing the layout selected in Aim 1. Eight participants will wear the test patches alongside conventional electrodes during resting tasks and a nap, using a tethered laboratory amplifier. Signals will be compared with the conventional electrodes using the Aim 1 criteria, and electrode–skin impedance will be tracked across the session.

A material combination is carried forward if impedance at 10 Hz increases by less than 50% from baseline over 8 h of wear, at least 90% of the sample remains adherent at removal, no reaction exceeds minimal erythema, comfort is not inferior to standard PSG electrodes, and test-patch signals meet the Aim 1 EEG and ECG criteria. If dry electrodes show excessive impedance, a thin hydrogel interface will be adopted; if adhesion is insufficient, a breathable overlay will be evaluated.

A shared processing pipeline will be developed for all recordings in this phase. It will include band-pass and notch filtering; automated signal-quality indices for each signal and epoch; detection of motion artifacts and contact loss; and separation of cervical EMG from EEG using spectral and learned methods suited to few-channel recordings. Outputs will include the comparison metrics in Table 1, computed reproducibly for every layout and material tested. The pipeline will be released as open-source code.

Public PSG datasets (Kemp et al., 2000; Goldberger et al., 2000; Zhang et al., 2018) will be used to train models restricted to signals that approximate what the patch could record: a single EEG derivation near the mastoid, ECG, and an EMG-derived muscle-tone signal. Features will include EEG band power and spindle and slow-wave measures; heart rate, RMSSD, and HF and LF power in windows centered on each 30-s epoch; and EMG power as an index of muscle tone. Two model classes will be compared: a feature-based gradient-boosted classifier and a deep sequence model (Supratak et al., 2017; Perslev et al., 2021). Ablation analyses will compare EEG-only, cardiac-only, and combined models, with particular attention to REM sleep.

Models from Aim 3b will be applied to the nap recordings from Aims 1b and 2c, with and without fine-tuning, to estimate how well they transfer to signals recorded at the patch site. Because afternoon naps contain little REM sleep, REM-specific evaluation in this phase relies on the public datasets, and transfer analyses focus on wake, N1, and N2.

All evaluation will use subject-wise cross-validation so that no participant contributes data to both training and testing. Outcomes will include epoch-by-epoch Cohen’s kappa relative to expert scoring, interpreted using conventional benchmarks (Landis and Koch, 1977); per-stage sensitivity, specificity, and F1 score, with emphasis on REM sleep; and the change in performance from adding cardiac and muscle-tone features to EEG.

Public PSG signals only approximate patch signals; this will be addressed by adding realistic noise and artifact profiles measured in Aims 1 and 2 to the public data. Class imbalance, particularly for N1, will be addressed with class-weighted loss functions and per-stage reporting. If five-stage performance is limited, a four-class scheme (wake, light NREM, deep NREM, REM) commonly used for wearable devices will also be reported.