A soft wearable near-infrared spectroscopy system for detecting brain water dynamics linked to glymphatic activity during sleep

Abstract

Sleep is a critical physiological process essential for overall brain health. Recent research shows that the glymphatic system is a key player in facilitating the metabolic waste clearance. However, the continuous real-time monitoring of brain water dynamics during sleep remains substantially challenging. Here, we introduce a soft, wearable near-infrared spectroscopy (NIRS) system to detect brain water dynamics potentially linked to glymphatic activity. The device features an integration of multi-wavelength LEDs and photodetectors. The in vivo study with multiple human subjects captures the device’s overnight sleep monitoring in a natural home environment, revealing continuous changes in brain water dynamics across different sleep stages. Our results support the link between sleep stage-dependent water dynamics and glymphatic activity. Spectral analysis identifies several physiological rhythms during sleep, including respiration, heart rate, and oscillations linked to slow-wave activity. These advancements significantly enhance the NIRS system’s potential for a deeper understanding of brain health.

A soft, wearable NIRS device monitors brain water dynamics linked to glymphatic activity during sleep.

INTRODUCTION

Sleep is essential for memory processing, cognitive function, and neural recovery (1–3). Disrupted sleep leads to the accumulation of metabolic waste, which interferes with cognitive function and memory formation (4, 5). The glymphatic system has been identified as playing a crucial role in cerebrospinal fluid (CSF) circulation, which involves cleaning metabolic waste and regulating central nervous system activity during deep sleep (6–8). Thus, monitoring brain water dynamics that reflect CSF redistribution provides an important indirect biomarker for understanding the relationship between sleep disturbances and neurological or systemic diseases (9). However, direct tracking of CSF flow remains a substantial challenge. Current invasive methods, such as two-photon microscopy, have limited clinical applications due to safety concerns, and Magnetic Resonance Imaging (MRI) remains constrained by the risk of intrathecal gadolinium injection, its high cost, the need for subject immobilization, and incompatibility with natural sleep environments (10, 11). In addition, polysomnography (PSG), the clinical gold standard for sleep assessment, is bulky, expensive, and is only available in hospital settings, making it unsuitable for long-term at-home monitoring (12). Table S1 summarizes the limitations of current devices for CSF/water-fraction detection.

Here, our study develops an all-in-one, soft, wireless, skin-conformal brain-water near-infrared spectroscopy (NIRS) system. This device ensures a high signal-to-noise ratio, stable and long-term monitoring with superior wearability during sleep. We hypothesize that CSF dynamics are reflected in brain-water dynamics measured at the forehead. In this approach, we utilize three wavelengths to extract the oxygenated hemoglobin (HbO), deoxygenated hemoglobin (HbR), and water fractions, enabling quantification of transient changes in brain-water content. The wearable device simultaneously captures physiological rhythms, such as respiration, heartbeat, and slow oscillation, closely related to brain water circulation. Beyond previous research, our device integrates these multi-scale features with wearable sleep staging, enabling comprehensive physiological assessment. Therefore, the purpose of this study is to demonstrate that brain water dynamics and associated physiological signals across sleep stages can be continuously, noninvasively, and physiologically meaningfully monitored using a soft NIRS system. Beyond simple sleep stage classification, this approach establishes the soft device as a platform for home-based, multi-scale physiological monitoring of brain fluid dynamics, potentially linked to clearance-related processes.

RESULTS

Overview of the soft wearable NIRS system for at-home detection of brain water dynamics related to glymphatic activity

Recently, MRI has been utilized for the study of the glymphatic system, but it is limited by cost and the need for natural sleep environments (8). To overcome these limitations, we developed a soft, wireless, wearable NIRS system, which enables continuous, real-time, home-based monitoring of brain water dynamics during sleep. Figure 1A presents an overview of the at-home monitoring NIRS system that measures brain water dynamics in both awake and sleep states. The soft device’s performance has been validated against simultaneously acquired electroencephalogram (EEG) and electrooculogram (EOG) signals from a commercial device, enabling direct correlation analysis between sleep stage classification and brain water changes. Figure 1B highlights the functional role of CSF in the glymphatic system. CSF removes metabolic waste products such as amyloid-β during deep sleep. Because direct measurement of brain-wide CSF dynamics is not feasible, we aimed to capture CSF-related physiological changes associated with glymphatic activity by tracking state-dependent variations in brain water content measured at the forehead. Figure 1C captures the mechanical and material properties of the soft NIRS device that can be easily attached to the forehead. The device utilizes light, flexible, and skin-tight polymers to ensure user comfort. It integrates LEDs emitting light at 640 nm, 680 nm, and 950 nm wavelengths with a photodetector that can detect multiple wavelengths to obtain various optical signals. LEDs mounted on the flexible circuit allow the device to bend with the skin, ensuring stable data acquisition under motion and prolonged use. Previous studies have largely relied on wired and invasive approaches during awake tasks (13), restrictive in-lab sleep studies (14–16), or stationary laboratory settings (8, 17, 18). On the other hand, our platform is wearable and operates wirelessly, enabling at-home natural sleep monitoring (Fig. 1D). To our knowledge, this is the first time a wearable device has detected brain water dynamics during at-home sleep conditions. Figure 1E shows the all-in-one wearable NIRS system that measures various signals and uses advanced AI-based analysis. The recorded EEG signals are classified into a hybrid model that combines threshold-based detection and AI machine learning algorithms. This allows the precise classification of sleep stages (Wake, NREM, REM) as well as the capture of various physiological rhythms such as respiration, heartbeat, and slow oscillation from the soft device at the same time. Brain water is quantified through the Beer-Lambert law. This multimodal analysis not only enhances the accuracy of sleep stage classification but also facilitates the identification of broader physiological mechanisms that influence sleep quality.

Fig. 1. Soft wearable NIRS system for at-home detection of brain water dynamics related to glymphatic activity.

(A) Sleep stage-dependent EEG signals and water fraction dynamics. (B) Glymphatic circulation illustrating CSF–ISF exchange and waste clearance during sleep. (C) Photo of a soft wearable NIRS system that is wireless, flexible, and low profile, including an integrated miniaturized battery embedded within the soft structure. (D) Significance of this work compared to other prior published articles in brain water assessment [wired & invasive during awake study: (13), restrictive in-lab sleep studies: (14–16), stationary settings studies: (8, 17, 18)]. (E) Flow chart showing how the wearable NIRS system can measure water dynamics during sleep, which are analyzed via AI-based machine learning and classification for sleep staging, physiological condition detection, and sleep quality assessment.

Overview of the device design and characterization of mechanical stability, performance, and thermal safety

As shown in Fig. 2A, the soft device comprises three main layers: an adhesive layer for skin attachment, a flexible printed circuit board (fPCB) that contains the LEDs and detector, and a silicone encapsulation layer for mechanical protection and comfort. The fPCB integrates multiple LEDs emitting at 640 nm, 680 nm (OIS-330 IE680, Opto International, Tokyo, Japan), and 950 nm (MTE9730CP Marktech Optoelectronics) wavelengths, and a photodetector (AS7341, ams OSRAM AG) is integrated to collect reflective signals. This photodetector can be acquired by simultaneously dividing signals of various wavelengths, as shown in fig. S1, allowing for precise tracking of multiple physiological parameters, including changes in brain water. Additionally, key electronic components, including Bluetooth low-energy (BLE) microcontrollers, voltage regulators, battery, and antennas, are integrated, enabling wireless communication and efficient signal processing. The device is powered by an onboard rechargeable Li-polymer battery (110 mAh, 3.7 V). Under the operating protocol used in this study, including continuous LED illumination and wireless data streaming, the system consumed an average power of approximately 70–75 mW, enabling ~5.5 hours of continuous operation under the experimental configuration. Recording duration is expected to be extendable in future designs through power optimization strategies, including duty-cycled LED operation, optimized firmware, and the use of higher-capacity batteries. Figire 2B shows the manufactured NIRS device, which features a miniaturized and skin-friendly design. LEDs and photodetectors are located on the device’s surface, allowing for the simultaneous detection of multi-wavelength signals, including near-infrared emission, during the sleep monitoring process. The finite element analysis (Fig. 2C and table S2) was further extended to evaluate the device mechanics under multiple bending curvatures (κ = 0.02 − 0.05 mm⁻¹), corresponding to radii of R = 50, 35, 25, and 20 mm. This range was selected to encompass typical adult forehead geometry as well as smaller infant-scale radii, thereby representing a more conservative (harsher) bending environment (19). Across all curvature conditions, the first Cu layer exhibited a maximum principal strain of 2.89 × 10⁻³ (0.29%), while the second Cu layer showed a lower strain of 1.18 × 10⁻³ (0.12%). The corresponding von Mises stress in the Cu traces remained below 1.07 × 10² MPa under the infant-scale bending condition (R = 25 mm), which is lower than the Cu yield strength (∼120 MPa). Even in the worst-case condition (R = 20 mm). Both strain and stress remained near but below the yielding threshold. These results indicate that the multilayer Cu interconnects are unlikely to undergo yielding or mechanical damage under physiologically relevant bending, supporting the mechanical robustness of the PCB. Figure 2D depicts the communication process in the soft NIRS device. The device collects physiological signals by emitting near-infrared light through multiple LEDs and detecting the reflected signals with an integrated photodetector placed on the skin surface. The acquired optical signals, containing information on brain water dynamics, are handled by the onboard BLE microcontroller, which performs signal digitization and data buffering. The converted data are wirelessly transmitted to an external computer via BLE, where further data processing, storage, and analysis are conducted. As shown in Fig. 2E, the signal-to-noise ratio (SNR) has been evaluated under sitting, walking, and climbing conditions. The rigid one exhibits low performance, with average SNR values of 6.42 dB (sitting), 2.38 dB (walking), and 2.38 dB (climbing). In contrast, the flexible device shows significant improvement, with SNR values of 12.72 dB, 7.44 dB, and 9.5 dB, respectively. The bar graphs highlight the superior signal stability of the flexible circuit across different motion conditions. The detailed SNR values, along with photos of rigid and flexible circuits, are provided in fig. S2.

Fig. 2. Overview of the device design and characterization of mechanical stability, performance, and thermal safety.

(A) Device layout including adhesive layer, fPCB, LEDs (640, 680, 950 nm), photodetector, BLE microcontroller, voltage regulators, battery, antenna, and Ecoflex encapsulation. (B) Photograph of the fabricated device. (C) Finite element simulation of multilayer strain under bending. (D) Schematic of wireless communication and signal processing. (E) SNR comparison between rigid PCB and flexible PCB (fPCB), showing superior stability of the fPCB during sitting, walking, and climbing. (F) Thermal safety validation: Pennes model simulation and FLIR measurements confirm that surface temperature remains below 41°C during 60 min of continuous operation.

Computational modeling study and experimental validation verify the device’s thermal safety during long-term operation of the soft NIRS system (Fig. 2F). Bioheat simulations indicate that skin-surface temperatures remain below the clinically acceptable 41°C safety threshold, with limited thermal penetration across tissue layers and stable temperature saturation over extended simulated operation (fig. S3). Extended simulations up to 300 minutes confirmed that no cumulative thermal drift occurred beyond the initial equilibration phase. The temperature distribution showed a clear spatial gradient, with the highest values at the skin surface and progressive lower temperatures toward deeper tissue layers, indicating that the optical heating was confined to the surface layer and did not cause excessive heat loads in the deep tissue. After 60 minutes of continuous operation, the 3D thermal distribution confirmed safe temperature levels across layers. Finite element simulations revealed that the tissue phantom reached thermal equilibrium after ~20 minutes of irradiation, with the maximum surface temperature stabilizing at 40.1°C, remaining below the 41°C safety threshold. The three-dimensional temperature field demonstrated a gradual decline from the skin surface toward deeper regions, decreasing to 38.5–39°C at the skin–skull interface and further to 36–37°C at the bottom of the skull layer. This gradient further supports that the applied optical heating was effectively confined to the superficial layers without inducing excessive deep-tissue thermal loading. The equilibrium temperature was maintained throughout the simulated duration (fig. S3). The initial heating rate was rapid (~0.7°C/min during the first 5 minutes), enabling the skin temperature to approach steady state within a short time scale. These computational findings provide quantitative evidence that the system can sustain long-term operation (300 min) within physiologically safe temperature limits (table S3). Experimental thermal imaging measurements exhibited the same stable heating pattern, further validating the computational predictions (fig. S4). These measurements also illustrate the skin temperature monitoring process during device operation, showing sequential tracking of surface temperature over time at different wavelengths.

Performance validation of the soft NIRS system for brain water monitoring

The working principle of the glymphatic system is illustrated in Fig. 3A. Additional details of the CSF pathways, astrocytic aquaporin-4 (AQP4) channels, and CSF–interstitial fluid (ISF) exchange are summarized in Fig. 3B, which highlights the mechanisms underlying glymphatic circulation. In this pathway, CSF enters along the perivascular spaces surrounding arteries, moves via AQP4 channels into the interstitial space, where exchanges with ISF occurs, and is subsequently directed toward the venous side, facilitating the removal of metabolic waste products (20). This clearance mechanism is essential for maintaining brain homeostasis and preventing neurological disorders. While CSF dynamics can be partially assessed using MRI, such approaches are not suited for continuous monitoring during sleep. As shown in Fig. 3C, the soft NIRS device was used to track state-dependent changes in brain water, providing a non-invasive readout of physiological processes associated with glymphatic activity. In this section, we validate the ability of the NIRS configuration to capture brain water dynamics within cortical tissue, which constitutes the primary physiological signal targeted by the present measurements. To assess the plausibility of this cortical depth sensitivity, we performed Monte Carlo simulations of photon propagation through layered head tissues (Fig. 3D) (21, 22). The photon density heatmap revealed that the greatest concentration of photon events occurred in superficial tissues, as expected given the high scattering coefficient of skin and skull. Importantly photon trajectories were not confined to superficial layers but extended beyond the subarachnoid CSF layer into deeper brain tissue. Layer tallies (Fig. 3E) further confirmed that although sensitivity decreased with depth, a measurable fraction of detected photons sampled cortical gray matter beneath the CSF layer. For the 3 cm source–detector separation used in this study, Monte Carlo simulations demonstrate depth-dependent sensitivity extending into cortical brain tissue. This validated the design choice of our source–detector spacing, showing that the optical geometry enables sampling of brain tissue depths relevant for cortical water measurements. These simulations underscore two crucial points. First, while superficial tissues dominate photon paths, they do not completely overwhelm the deeper signal; instead, depth penetration into cortical tissue is consistently preserved. Second, because cortical brain tissue contains a high water fraction with distinct optical absorption characteristics, combined photon sampling across multiple channels can produce detectable changes in Δ[H₂O] reflecting cortical brain water dynamics. Overall, the simulations provided a mechanistic basis for the feasibility of our approach by demonstrating that the NIRS configuration is well suited for tracking state-dependent changes in cortical brain water, which informed the interpretation of subsequent in vivo measurements.

Fig. 3. Performance validation of the soft NIRS system for brain water monitoring.

(A) Schematic illustration of glymphatic circulation and CSF–ISF exchange. (B) Detailed glymphatic system showing CSF influx, exchange via AQP4 channels, and waste clearance through perivenous routes. (C) Three LEDs and their wavelengths, and the inter-distance between LEDs and the detector on the forehead-mounted device. (D) Monte Carlo simulation of photon propagation through the tissues on the forehead. (E) Layer-wise photon distribution across tissue depth, demonstrating sensitivity to brain tissue beneath superficial layers using the soft wearable NIRS system. (F) Optical intensity changes at 640, 680, and 950 nm during daily activities, including repeated motions and rests. (G to J) Zoom-in data of Δ[H₂O] capturing the increasing trends during active motions, while decreasing behavior during rests. (K) Biphasic Δ[H₂O] response during cyclic breath-holding and relaxing.

An additional set of tests shows the measured water dynamics during alternating cycles of physical activity and rest (Fig. 3F). Across repeated transitions, the recorded optical intensities at 950, 680, and 640 nm are shifted in magnitude in ways that aligned with state transitions. From these signals, we derived a composite metric, Δ[H₂O], using the modified Beer–Lambert law, optimized for sensitivity to water-linked changes. Across multiple workout/rest cycles, Δ[H₂O] exhibits reproducible biphasic modulations: during activity, Δ[H₂O] increases initially and then declines midway through the epoch, while during rest, Δ[H₂O] decreases first and then rises again toward the midpoint (Fig. 3, G to J). The early rise in Δ[H₂O] during activity likely reflects the physiological increase in metabolic demand, which elevates cerebral and peripheral blood flow and delivers greater plasma water to tissues, thereby enhancing water-linked absorption signals (23). The subsequent decline may correspond to autoregulatory adjustments in blood and CSF distribution as metabolic needs stabilize. Conversely, the initial drop during rest likely reflects the reduced vascular tone and perfusion, followed by a rebound increase as interstitial fluids and CSF re-equilibrate. These alternating patterns underscore that brain water dynamics are not strictly monotonic but instead capture the interplay between rapid hemodynamic shifts and slower fluid redistribution processes. We also tested the device in subjects during alternating cycles of normal breathing and voluntary breath holds (Fig. 3K). These states are associated with predictable physiological shifts: breath holding elevates CO₂ and intracranial pressure, whereas normal breathing stabilizes these variables (24, 25). Correspondingly, Δ[H₂O] reliably distinguishes between the two conditions, exhibiting a characteristic biphasic response during breath holds—an initial increase, followed by a gradual decrease, before returning toward baseline (26). This pattern likely reflects the combined effects of hypercapnia-driven vasodilation, which elevates cerebral blood volume and water-linked absorption, and subsequent regulatory adjustments as intracranial pressure induces broader intracranial fluid shifts, including CSF redistribution. In this context, the breath-holding task primarily validates the sensitivity of the device to cerebrovascular and composite fluid-dynamic responses, rather than directly validating CSF-specific dynamics. During normal breathing, by contrast, Δ[H₂O] shows a slight but consistent increase, reflecting stable ventilation, clearance of CO₂, and maintenance of vascular tone. These state-dependent shifts were reproducible across repeated cycles and did not require recalibration of the device. This stability shows that the system is robust enough for prolonged monitoring, a prerequisite for sleep studies. Taken together, the Monte Carlo simulations, activity/rest validations, and respiratory validation show a consistent conclusion that Δ[H₂O] reflects physiologically meaningful brain water dynamics, detectable noninvasively with our wearable system. Comparison with other biosignals (fig. S5) indicates that the observed low-frequency components are not dominated by cardiac or peripheral pulsatile signals (PPG/HR-related artifacts). In addition, cross-talk analysis between HbO, HbR, and H₂O (table S4) confirms that Δ[H₂O] is weakly correlated with hemoglobin-related signals, further supporting the stability of Δ[H₂O] estimation against hemoglobin cross-talk and physiological noise. These results motivated us to investigate further the coupling of brain water dynamics with sleep stages, as detailed in the following section.

Sleep stage determination and brain water dynamics assessment

The combined configuration of the EEG, EOG, and soft NIRS monitoring system is illustrated in Fig. 4A. Figure 4B shows continuous, long-term EEG and NIRS recordings, demonstrating stable tracking of brain-water–related optical changes alongside neural activity. After establishing the reliability of Δ[H₂O], we evaluated its relationship to sleep. For this purpose, we combined EEG and EOG with the NIRS device for overnight recordings. This multimodal design allowed us to compare traditional neural markers of sleep with NIRS-derived brain water dynamics. We then implemented and compared three approaches for sleep-stage classification: (1) a machine-learning model, (2) a threshold-based model, and (3) a hybrid model. All methods were applied to 30-second non-overlapping segments of the preprocessed EEG and EOG signals. Figure S6 illustrates hypnograms derived from two expert scorers, our machine learning model, and a hybrid model for sleep stage evaluation. As shown in Fig. 4C, the feature extraction framework categorizes EEG and EOG signals into multi-domain feature sets for sleep stage characterization. Features are organized into three categories. Spectral features compute theta band power (4–8 Hz) present during drowsiness and N1 sleep onset, alpha band power (8–13 Hz) prominent during relaxed wakefulness and attenuated during sleep, sigma band power (11–15 Hz) for sleep spindle detection characteristic of N2 sleep, beta band power (13–35 Hz) associated with active wakefulness and cognitive processing, gamma band power (35–45 Hz) reflecting high-frequency cortical activity, and alpha-to-theta power ratios that decrease during sleep onset. Hjorth features calculate activity for signal variance, mobility for normalized frequency dispersion, and complexity for bandwidth measurement (27). Nonlinear features compute sample entropy for signal regularity quantification and phase-amplitude coupling, measuring high-frequency amplitude modulation by low-frequency phase (28, 29). EOG features are categorized into spectral, temporal, and statistical domains. Spectral features analyze unfiltered total power, low-frequency power (0.5–4 Hz) for slow eye movements during sleep onset and transitions, and mid-high frequency power (4–15 Hz) for rapid eye movements during REM sleep across both EOG channels. Temporal features include Saccades for rapid eye movement event detection and Slope changes through derivative zero-crossing counts for movement direction changes. Statistical features compute the Mean for baseline signal levels, the Variance for amplitude variability, the Skewness for distribution asymmetry, and the Kurtosis for extreme event detection.

Fig. 4. Sleep stage determination and brain water dynamics assessment.

(A) System overview showing the soft NIRS system for measuring brain water dynamics and a commercial device that measures EEG and EOG signals. Both devices measure signals while a subject sleeps at home. (B) Representative overnight recorded data showing EEG signals and optical signal with 3 LEDs (640/680/950 nm). (C) Feature groups extracted from EEG and EOG signals for automated staging. (D to E) Staging performance comparison of our method versus traditional ML and deep-learning methods. (F) Rule-based threshold pipelines: relative σ-power (10–15 Hz) with a moving-average threshold for NREM/REM. (G) Example of band power analysis of σ-power and adaptive threshold. (H) Confusion matrix of sleep stage classification for the threshold model. (I) Hybrid workflow that combines threshold data with an ML classifier. (J) Accuracy distribution across context windows for the hybrid model. (K) Confusion matrix for the hybrid model showing improved accuracy and balance between sleep stages. (L) Overnight hypnogram produced by the hybrid classifier. (M) Multitaper spectrogram of EEG (upper panel) and NIRS-derived brain water signal (lower panel) across the night. (N) Power spectral analysis of the NIRS-derived brain water signal, highlighting distinct peaks corresponding to low-frequency brain water oscillation, respiratory signal (RS), slow-oscillation–linked NIRS oscillation (SONO), and cardiac-cycle (CC). (O) Boxplots of transition timing reveal a near-zero lag between Δ[H₂O] state boundaries and EEG-defined stages, indicating a tight coupling of brain-water dynamics with sleep architecture.

As summarized in Fig. 4, D and E, the performance evaluation presents comparative results across traditional machine learning and deep learning models. Traditional machine learning algorithms recorded validation accuracies of 0.941 for RandomForest, 0.931 for GradientBoosting, 0.921 for LightGBM, and 0.917 for XGBoost. Test accuracies of 0.846, 0.800, 0.773, and 0.756, respectively, resulting in validation-test accuracy differences of 0.095, 0.131, 0.148, and 0.161. Deep learning architectures yield validation accuracies of 0.901 for MLP, 0.897 for CNN, and 0.884 for LSTM, with corresponding test accuracies of 0.870, 0.850, and 0.820, producing validation-test differences of 0.031, 0.047, and 0.064. The proposed Multi-window CNN-LSTM architecture shows validation accuracy of 0.910 and test accuracy of 0.870, yielding a validation-test difference of 0.040. The proposed approach maintains consistent performance across evaluation metrics with reduced validation-test performance differential relative to traditional methods, indicating improved generalization capacity for cross-dataset evaluation scenarios. As a baseline, we built a sigma-band-based sleep staging algorithm that used a moving-average–based threshold to separate Wake, NREM, and REM states (Fig. 4F). First, we calculated sigma-band power (10–15 Hz) and total power from the EEG channels. The sigma-band power was compared against a dynamic threshold equal to 60% of its moving average. If sigma power was above this level, the system labeled the epoch as NREM; if it was below, it was labeled as REM. At the same time, we checked the total power against a fixed cutoff of 50%. If total power went above that cutoff, the system overrode the sigma decision and classified the epoch as Wake. In this way, sigma activity managed the NREM versus REM distinction, while high overall power signaled Wake. This two-step method was simple to run and efficient. In practice, this method shows that NREM detection was strong (95.9%), but REM detection was weaker (about 70%). As a baseline, we implemented a moving average threshold sigma-band power detector (Fig. 4G). This threshold-based method segmented Wake, NREM, and REM states with moderate success. NREM detection was particularly strong (95.9%), reflecting the reliability of sigma activity as a spindle marker. REM detection, however, lagged at ~70%, underscoring the limitations of single-feature approaches (Fig. 4H). The threshold method nonetheless established a proof-of-concept, showing that sleep staging could be approximated in real time using computationally light algorithms. To overcome the threshold model’s limitations, we developed a hybrid system combining sigma-based thresholds with a supervised classifier trained on the full feature set (Fig. 4I). This hybrid pipeline exploited the strengths of both approaches: thresholding provided physiologically grounded priors, while machine learning corrected misclassifications and sharpened transitions. The result was a significant accuracy improvement, with overall classification rising to 80–90% as shown in Fig. 4J. The model particularly boosted REM detection, which is traditionally the most challenging stage to classify (Fig. 4K). When compared against conventional ML models and deep learning baselines, the hybrid classifier consistently outperformed in terms of accuracy, interpretability, and efficiency. The hybrid approach, therefore, represents an optimal balance for wearable deployment: high accuracy with low cost and interpretable decision boundaries. Figure 4L shows sleep stage classification based on EEG signals from the commercial device, recorded concurrently with the soft NIRS device. Most importantly, Δ[H₂O] exhibits state-dependent coupling with EEG-defined stages. Multi-taper spectrograms (Fig. 4M) reveal oscillatory patterns in Δ[H₂O] that were temporally aligned with EEG-based sleep-stage dynamics. At the same time, we found several distinct spectral peaks in Fig. 4N. These peaks were classified according to their frequency ranges, including low-frequency brain water oscillation (~0.05 Hz), respiratory-related activity (RS; ~0.3 Hz), slow oscillation–range NIRS oscillation (SONO; 0.6–0.7 Hz), and cardiac cycle (CC; ~0.8–1.2 Hz). The frequency bands of RS and CC are in line with previous fNIRS studies (8, 30–32). In terms of SONO, the observed spectral peak around 0.6 to 0.7 Hz overlaps with the EEG slow oscillation (~0.5 Hz) (33, 34), suggesting a potential association between SONO and electrophysiological activity. Further studies incorporating direct reference signals (e.g., simultaneous PSG-derived oscillatory metrics) are required to validate the physiological specificity of these spectral components and to refine the frequency-band boundaries associated with each process. Together, these results indicate that NIRS-derived brain water signals exhibit structured, state-dependent spectral components that are temporally coordinated with EEG-defined sleep architecture, a pattern consistent with sleep-linked brain fluid dynamics previously implicated in glymphatic activity. Having established that brain water dynamics vary systematically with EEG-defined sleep stages, we next examined whether these dynamics are temporally aligned with sleep-stage transitions. To quantitatively assess temporal coupling between brain-water dynamics and sleep-stage transitions, we performed a transition-based lag analysis. For each sleep-stage transition, the temporal lag was defined as the difference between the EEG/EOG-defined stage boundary and the nearest transition point detected in the Δ[H₂O] signal (Δt = tΔ[H₂O] − tEEG). These lag values were computed for all transitions across subjects and nights and pooled to generate lag distributions. As shown in Fig. 4O, median lags were −1.45 s during wakefulness, 0 s during REM sleep, and 0 s during NREM sleep. The corresponding interquartile ranges were −6.1 to 2.0 s (Awake), −5.7 to 4.5 s (REM), and −4.3 to 3.15 s (NREM). These distributions are centered near zero across states, indicating minimal temporal offset between Δ[H₂O] transitions and EEG-defined sleep-stage boundaries and demonstrating tight temporal alignment between NIRS-derived brain water dynamics and neural state transitions. This near-zero lag reflects rapid coordination between brain water dynamics and neural state transitions, rather than a delayed or secondary process.

Analysis of sleep stage-dependent brain-water dynamics and physiological coupling during at-home sleep

Unlike short-term laboratory tests such as breath holding or exercise cycles, overnight recording imposes greater demands on device stability, subject comfort, and signal robustness. Fig. 5, A to E present a single-night recording from one participant to demonstrate the temporal relationship between sleep stages and brain water dynamics. Figure 5A illustrates that the wearable device sustains stable, high-quality measurements throughout a full night of sleep (movie S1). In post-recording analysis, sleep stages classified from EEG/EOG signals were aligned with the Δ[H₂O] signal, which captures band-limited spectral features corresponding to RS, SONO, and CC. SONO is a feature of NREM sleep, and its power is elevated during deeper sleep states (35, 36). In NREM, physiological spectral peaks appear narrower and higher, whereas in REM, they are broader and slightly lower in amplitude. Figure 5A shows time-resolved band-limited power within each target frequency band, whereas Fig. 5, B to D specifically focus on power spectral density (PSD)-derived peak power within the same bands. The band-limited power can occasionally appear higher during REM in representative segments, while the spectral peak power is higher during NREM, reflecting broader spectral distributions during REM and more concentrated peaks during NREM. The mean SONO peak power derived from Δ[H₂O] signals is lowest during wakefulness, increases slightly during REM, and rises during NREM (Fig. 5B). The NREM sleep stage is well known for slow and synchronized neural activity, which has been associated with changes in brain fluid dynamics in prior studies compared to wakefulness (37). During REM sleep, a paradoxical sleep stage that more resembles wakefulness rather than NREM, neural activity becomes relatively faster and desynchronized, accompanied by distinct changes in brain water dynamics (38, 39). The changes across sleep-stage transitions from Wake to REM and NREM were also demonstrated by Δ[H₂O] measurements, suggesting that water dynamics can serve as a complementary, non-invasive biomarker for sleep-stage transitions. RS dynamics shows clear state-specific differences in the PSD (Fig. 5C). During REM sleep, the RS peak was centered at approximately 0.35 Hz, had lower power, and was broadened—features that indicate irregular and variable breathing, consistent with prior reports of respiratory rate variability in healthy sleeping adults (40–44). By contrast, during NREM, RS slowed to ~0.25 Hz, the peak narrowed, and power increased, reflecting steadier, more powerful breathing. These findings align with prior polysomnography work showing that REM is marked by erratic RS, whereas NREM features more regular breathing patterns. These differences are detectable through Δ[H₂O], suggesting that brain water-sensitive optical dynamics may reflect sleep-stage dependent cardiorespiratory coupling (Fig. 5D) (45–48). In NREM, the 0.9 Hz peak was high in power and narrow in width, Indicating a more regular cardiac-frequency-range spectral feature. (49). As subjects transitioned into REM, this peak broadened and reduced in power, reflecting increased variability in cardiac-frequency-range dynamics. Importantly, these changes occurred in tight synchrony with EEG-defined transitions, showing the physiological sensitivity of Δ[H₂O]. These observations should be interpreted cautiously. Despite prior reports linking similar frequency bands to physiologic processes, the acquired RS, SONO, and CC signals are spectral features derived from Δ[H₂O], which were not quantitatively validated against concurrent reference signals or gold-standard PSG. To better visualize how these physiological rhythms separate sleep states, we projected features for slow oscillations, RS, and CC into a joint feature space (Fig. 5E). This visualization suggested partially separable clusters of epochs: NREM and REM tended to occupy distinct regions in the SONO–RS–CC feature space, with some overlap (50, 51). The boundary between clusters was confined to a narrow margin, indicating potential separability of sleep states at the epoch level. However, this result is presented as an illustrative visualization rather than quantitative demonstration of sleep-stage classification based on Δ[H₂O] alone. Nonetheless, these feature-space patterns suggest the feasibility of wearable optical devices potentially supporting basic sleep monitoring as a complementary modality. At the group level, we next examined cumulative brain water trajectories across sleep-stage transitions. Figure 5, F to I were generated from multiple nights across participants (n = 4; total 16 nights) and summarize the mean trajectory with the standard error. These results illustrate that (i) entry into NREM from both Wake and REM was accompanied by rising accumulated brain water, which continued to increase through early and sustained NREM. (ii) In contrast, transitions from NREM into REM were marked by decreases in accumulated brain water (fig. S7). These direction-specific patterns suggest that brain water dynamics are dependent on sleep-stage transitions. Such patterns are consistent with prior models proposing sleep-stage dependent regulation of brain fluid dynamics, with NREM sleep associated with greater fluid accumulation relative to REM sleep (8, 16, 38, 39).NREM sleep was consistently associated with higher Δ[H₂O], stronger slow oscillations, and more powerful RS and CC rhythms. REM, in contrast, was characterized by reduced Δ[H₂O], irregular RS, and attenuated cardiac rhythms. Transitions between states were accompanied by direction-specific Δ[H₂O] shifts, suggesting that brain water regulation co-varies with the transitions themselves (52, 53).

Fig. 5. Analysis of sleep stage-dependent brain-water dynamics and physiological coupling during at-home sleep.

(A) Overnight sleep recording data showing hypnogram (W: Wake, N: NREM, R: REM), Δ[H₂O], RS (respiratory signal), SONO (slow oscillation–linked NIRS oscillation), and CC (cardiac cycle), well aligned with sleep-stage transitions. (B) Mean power of SONO across events, highest during NREM. (C) RS spectra showing narrower, higher ~0.25 Hz peaks in NREM and broader, lower ~0.35 Hz peaks in REM. (D) Three-dimensional feature space showing CC power and variation during transition from NREM to REM. (E) 3D feature-space visualization of Δ[H₂O]-derived metrics (SONO, RS, and CC) across epochs labeled as NREM and REM. Each axis corresponds to the SONO, RS, and CC feature values. No dimensionality reduction or clustering algorithm was applied. (F to I) Transition-dependent accumulated brain water changes: Wake→NREM (increase), NREM (increase), NREM→REM (decrease), and REM → NREM (increase). The accumulated trace is shown as a descriptive visualization of cumulative Δ[H₂O] within each temporal window. Grand-averaged accumulated brain-water trends are displayed as a solid line, with standard errors shown as shaded areas.

DISCUSSION

This paper presents a wireless, soft wearable NIRS device designed to monitor real-time, long-term brain water dynamics associated with transitions between sleep stages in a home setting. This at-home usable device offers insights into the physiological mechanisms underlying sleep. The low-profile design incorporates multi-wavelength LEDs and photodetectors with flexible circuits and soft encapsulants, ensuring skin-conformal contact. This design minimizes motion artifacts and enhances user comfort during sleep monitoring. In vivo studies involving multiple human subjects demonstrate the device’s capability for overnight sleep monitoring in a natural environment. During sleep, optical measurements can be influenced by potential confounds, including variations in breathing depth, subtle forehead pressure changes, and slow temperature drift. Therefore, this study focuses on sleep-stage–dependent trends rather than absolute water values. In this context, even in the absence of explicit correction for all motion- or pressure-related artifacts, movements and pressure-induced coupling changes may partially contribute to Δ[H₂O], particularly during long overnight recordings. The motion-robustness analyses should be interpreted as comparative validation of flexible versus rigid implementations, which is distinct from posture-specific motion and pressure effects encountered during natural sleep. The findings reveal continuous changes in brain water dynamics across different sleep stages, demonstrating robust sleep-stage–dependent modulation of brain water signals. Additionally, spectral analysis reveals Δ[H₂O]-derived spectral features during sleep within frequency ranges commonly associated with respiration, cardiac activity, and slow-oscillation–range dynamics. Although this study does not directly measure cerebrospinal fluid flow or solute clearance, the observed sleep-stage–dependent brain water dynamics occur within physiological regimes in which glymphatic activity is known to be modulated. Glymphatic function has been shown to vary across sleep stages, with enhanced activity during NREM sleep and attenuation during REM sleep, in parallel with changes in neural activity, vascular dynamics, and intracranial fluid regulation (4, 8, 16, 38, 39, 54, 55). In this context, the wearable NIRS platform’s ability to capture state-locked brain water dynamics with high temporal resolution during natural sleep suggests its potential utility as a noninvasive tool for monitoring glymphatic-related brain fluid dynamics longitudinally (13, 16, 56).

Collectively, this study demonstrates the feasibility of the soft, wearable NIRS device for capturing sleep-linked brain water dynamics in naturalistic settings. This approach has potential applications in longitudinal sleep monitoring and complementary assessment of sleep physiology, and may be particularly relevant for neurological conditions in which sleep and brain fluid dynamics are altered. Additionally, this device can be integrated into next-generation wearable healthcare platforms, thereby enhancing personalized, at-home healthcare solutions. Additional work will incorporate multi-site sensing and short-separation channels to improve spatial coverage and account for superficial contributions. Device designs will further improve mechanical compliance and skin conformity by separating the optical transmitter and receiver into smaller functional islands connected by compliant interconnects. In addition, short-separation channels will be incorporated to better account for superficial scalp signals. Future studies incorporating concurrent reference measurements, such as contrast-enhanced MRI or disease-specific biomarkers, will establish specificity and define the translational role of this approach in glymphatic research.

MATERIALS AND METHODS

Mechanical reliability analysis using finite element simulation

Finite element simulations were performed using ABAQUS (Dassault Systèmes Simulia Corp.) to evaluate the mechanical reliability of multilayer PCB structures during bending. A 3-point bending configuration was applied to forehead-like deformation, and the bending condition was set according to values based on bending parameters reported previously in the literature (19). The PCB model consisted of polyimide (PI) substrates and embedded copper (Cu) interconnects. Material properties were defined as: _EPI = 2.5 GPa, νPI = 0.34; and _ECu = 110 GPa, νCu = 0.34. The analysis focused on extracting strain distributions across the Cu layers.

Bioheat transfer modeling

Bioheat transfer was modeled in MATLAB using a one-dimensional Pennes’ framework across two layers (skin 3 mm, skull 7 mm), each discretized at 0.1 mm intervals. Layer parameters—density (ρ), specific heat (c), thermal conductivity (k), perfusion rate (ω), and metabolic heat generation (Q_met)—were drawn from published values. Optical heating (Q_opt) was computed by distributing a 5.6 mW LED source over depth using layer-specific absorption coefficients (μ_a) and beam geometry, following Beer–Lambert attenuation. The transient heat equation was solved explicitly with a 0.01 s time step for a 60 min simulation, holding deep tissue at core temperature and omitting surface convection to focus on subsurface heating. Outputs were rendered as depth–time temperature heat maps, delineating the interplay of conduction, perfusion, metabolism, and optical absorption. To assess wavelength dependence, the entire bioheat simulation was repeated separately for each NIRS wavelength (640 nm, 680 nm, 950 nm). For each run, μ_a values were set to the extinction coefficients specific to that wavelength while all other parameters remained constant. This yielded wavelength-specific temperature profiles, allowing direct comparison of heating distributions across the three optical channels. The numerical solution of Pennes’ bioheat equation was implemented as originally formulated (57) and incorporates layer-specific perfusion effects as described by Weinbaum et al. (58).

Thermal stability analysis using finite element simulation

Finite element simulations were performed in COMSOL Multiphysics (v6.2) to analyze heat transfer in a skin–skull phantom under LED irradiation. The model was a 10-mm cylinder consisting of 3 mm skin and 7 mm skull. Material properties, perfusion, and metabolic heat were assigned from table S5. Boundary conditions included convection and radiation (emissivity = 0.95, ambient = 25.0°C). The initial temperature was 36.5°C. Optical heating was modeled with a Gaussian LED beam (power = 5.6 mW, half-angle = 10°) following PD = P_LED/A, where A is the beam area at depth z., with absorption coefficients μ_a = 0.025 mm⁻¹/0.05 mm⁻¹ (950 nm/680 nm, skin) and 0.01 mm⁻¹/0.02 mm⁻¹ (950 nm/680 nm, skull). Transient simulations (0–300 min) yielded spatiotemporal temperature distributions, confirming that the temperature rise stabilized below clinically relevant safety limits. These results demonstrate the thermal stability of the system during continuous operation.

Monte Carlo model of light propagation

Monte Carlo photon propagation was simulated in MATLAB to map near-infrared light transport through a five-layer head model: skin (3 mm), skull (7 mm), CSF (2 mm), gray matter (4 mm), and white matter (10 mm). Each layer was assigned literature-based absorption (μa) and reduced scattering (μ′s) coefficients with anisotropy g = 0.90 (table S6). We launched N = 5 × 10⁵ photons with initial angles randomized ±10° from vertical. Photons propagated in steps of length s = −ln(r)/μt (where μt = μa + μs), underwent scattering via a 2D Henyey–Greenstein phase function, and had weights attenuated by μs/μt. Low-weight photons were terminated by Russian roulette. Photons exiting at z ≤ 0 within a 1 mm detector radius at a 30 mm lateral offset were recorded. Photon trajectories were histogrammed in 0.2 mm bins and Gaussian-smoothed to produce 2D “banana” heat maps.

Calculation of brain water change (Δ[H₂O] and cumulative brain water)

We calculated brain water changes using multi-wavelength NIRS at 640, 680, and 950 nm. Raw light intensity signals were converted to optical density changes relative to baseline. Using the modified Beer–Lambert law, we applied an extinction matrix for HbO, HbR, and H₂O and corrected for differential pathlength factors (DPFs) as shown in table S7. The use of three wavelengths ensured accurate separation of water fraction from hemoglobin signals, thereby enabling reliable quantification of CSF-associated brain water changes (13). Because the NIRS signals were acquired in reflectance mode from the forehead, CSF-associated brain water changes include contributions from superficial tissues such as the scalp and skull in addition to deeper compartments and therefore should not be interpreted as exclusively CSF-isolated measures. The resulting ΔOD (Optical Density) signals were transformed into concentration changes, with the H₂O channel used to track brain-water dynamics. Instantaneous brain-water changes were estimated as the derivative of the H₂O signal, reflecting moment-to-moment fluctuations in brain water. In contrast, accumulated brain-water changes were calculated as the time-integral (cumulative sum) of these instantaneous variations across windows of interest, providing the net displacement of brain water over the epoch. Cumulative brain water was computed as the cumulative sum of the baseline-normalized Δ[H₂O] signal derived from multi-wavelength NIRS (fig. S8). Δ[H₂O] is first normalized to a baseline period (the first 5 minutes of each analysis window), and integration begins at the window start, yielding the net brain-water displacement relative to baseline. To mitigate slow instrumental and physiological drifts, linear detrending and temporal smoothing are applied to the instantaneous Δ[H₂O] signal prior to integration, and the same processing is applied to the cumulative signal. Because the cumulative signal represents the time-integral of Δ[H₂O], monotonic increases or decreases reflect sustained directional shifts in brain-water dynamics rather than uncorrected signal drift. Here, this “accumulated” trace is used as a visualization aid for an intuitive presentation of the direction and persistence of Δ[H₂O] changes within the temporal window, rather than as a validated absolute measure of water accumulation or CSF dynamic. In this context, we did not consider specific signal processing approaches, including subject-level normalization. These preprocessing steps do not include explicit artifact correction or separation of potential harmonic components. Together, these measures quantified both the dynamic rate of change and the total shift in brain-water during non-rapid eye movement (NREM) and rapid eye movement (REM) sleep onset, allowing comparison of brain-water dynamics across sleep states.

Experiment for device validation (1) exercise

To validate the soft NIRS device, a trial was done (N = 4) with two back-to-back cycles of 20 s seated rest followed by 30 exercises, while continuous light intensities at 640 nm, 680 nm, and 950 nm were sampled at 10 Hz. For each rest and exercise epoch, the H₂O-derived concentration trace was first linearly detrended and smoothed with a 30 s moving average filter to remove drift and high-frequency noise. Instantaneous water change was quantified as the slope of Δ[H₂O] over each window—obtained via least squares regression—reflecting the rate of fluid shift. Cumulative change was calculated by integrating the Δ[H₂O] curve, yielding the total fluid displacement. This approach captures both rapid and aggregate CSF dynamics under controlled rest versus exercise conditions.

Experiment for device validation (2) breath-holding

To further validate the device, a trial was conducted (N = 4) with two back-to-back cycles of 60 seconds of breath-holding followed by 20 seconds of normal breathing. Continuous light intensities at 640 nm, 680 nm, and 950 nm were recorded at 10 Hz. The H₂O concentration trace was first linearly detrended and smoothed with a 30 s moving average filter to eliminate drift and high-frequency noise. Within each breath hold and recovery window, we quantified instantaneous water change as the slope of Δ[H₂O] via least squares regression and cumulative change by integrating the Δ[H₂O] curve. This approach allows us to compare CSF/brain water changes during different stages in the breath-holding/breathing cycle and validate the device functionality.

ML Data Preparation

Sleep data from 40 apnea patients, with 35,927 epochs, was obtained from the Institute of Systems and Robotics, University of Coimbra (ISRUC) public PSG dataset. Of the 35,927 epochs, 829 epochs contained apnea events, and 35,098 epochs were labeled normal. 32 healthy participants with 15,590 epochs. Of the 15,590 epochs, 1,883 were W, 685 were N1, 5,858 were N2, 4,404 were N3, and 2,760 were R. Four patients from separate sleep data from the ISURC public PSG dataset and four healthy participants’ data were used as the test data to evaluate the performance of the proposed methods. Artifact Subspace Reconstruction (ASR) was applied to remove transient artifacts including muscle activity and movement-related noise while preserving sleep-related neural patterns (59). To ensure consistency across heterogeneous datasets, statistical alignment was performed using scaling with high-pass filtering (0.5 Hz cutoff) to normalize the data distributions of healthy patients to match public PSG data characteristics while preserving signal integrity across channels. After alignment, signals were bandpass filtered from 0.3 to 30 Hz and notch filtered at 59–61 Hz. The bandpass range followed AASM guidelines, and the notch filter suppressed 60 Hz power line interference.

Sleep classification model: ML

Sleep staging was performed exclusively using EEG and EOG signals, and the Δ[H₂O] signal was analyzed only after sleep stages were determined and was not included as an input feature in any of the classification models. Table S8 summarizes the EEG/EOG features used for the sleep-stage classification models, including their categories, descriptions, and physiological relevance. A convolutional neural network (CNN) combined with bidirectional long short-term memory (bi-LSTM) was implemented for the classification of sleep stages. Input data consisted of synchronized electroencephalogram (EEG) and electrooculogram (EOG) recordings segmented into 30-second epochs. Each epoch was divided into six consecutive 5-second windows to capture temporal dynamics in sleep stage transitions. The EEG signals were processed as 2-channel inputs with sequence length 11, while EOG signals were processed as 2-channel inputs with sequence length 9, resulting in input shapes of (batch_size, 6, 11, 2) and (batch_size, 6, 9, 2), respectively. The CNN component employed two parallel encoders for modality-specific feature extraction. Each encoder comprised four 1D convolutional layers with filter progression of 128, 256, 128, and 64 (kernel size = 3, padding = 1). Batch normalization was applied after the 256-filter layer, followed by leaky ReLU activation throughout all convolutional blocks. Max pooling (kernel size = 2, stride = 2) was applied, followed by a fully connected layer with 256 units and dropout regularization (rate = 0.5). Each 5-second window was independently analyzed by both CNN encoders, producing 256-dimensional feature vectors per modality. EEG and EOG features were concatenated at the window level to form 512-dimensional representations. These six window features were stacked into a sequence of shape (batch_size, 6, 512) and fed into a 2-layer bidirectional LSTM with 256 hidden units per direction. The final LSTM output was passed through dropout (p = 0.5) and a linear classifier to predict one of the sleep stage classes. The model was optimized using Adam (learning rate = 3.5e-4, weight decay = 1e-5) with weighted cross-entropy loss. The learning rate was reduced by a factor of 0.5 at the validation accuracy plateau (patience = 5), and early stopping (patience = 20) prevented overfitting. Model hyperparameters were selected using random search strategies.

Sleep classification model: threshold

In parallel, we implemented a rule-based threshold model focused on the σ (spindle) band. For each epoch, we computed relative σ power (10–13 Hz) normalized by the total 0.5–30 Hz power. A dynamic threshold was set at 60% of the 75th-percentile σ power within a 25-epoch rolling window. Epochs with σ power above this threshold were labeled REM; below threshold, NREM (60). We enforced a minimum dwell time of four consecutive epochs to prevent rapid flips, and designated pre-sleep epochs (high total power) as Wake.

Sleep classification model: hybrid

To leverage the strengths of both approaches, we combined labels via a local voting scheme. For each epoch, we collected labels from a 5-epoch context window (2 before, current, two after) of both our classifier and the threshold model, resulting in 10 votes per epoch. The final label was chosen by majority vote. This hybrid method improved temporal consistency and mitigated isolated misclassifications, producing a smooth, physiologically plausible hypnogram.

Power spectrogram analysis

We began with the 10 Hz H₂O concentration time series, first applying a linear detrending step to remove baseline drifts arising from long-term physiological fluctuations or device thermal effects. Next, to isolate only the ultra-slow band (0.01–0.5 Hz) associated with glymphatic CSF pulsations, we implemented a zero-phase FIR band pass filter with an impulse response length of ~60 s, chosen to achieve a sharp cutoff while avoiding phase distortion. For time–frequency decomposition, we selected the multilayer method, which uses orthogonal Slepian tapers to minimize spectral leakage and reduce variance compared to single window approaches. We set the time–bandwidth product TW = 3 (yielding K = 2·TW–1 = 5 tapers), a balance that provides ≈0.05 Hz spectral resolution while maintaining stable power estimates in each window. The choice of five tapers follows prior electrophysiological studies of low-frequency oscillations, where it proved optimal for revealing slow rhythms without excessive smoothing. A 120-s sliding window with 10-s hop was used so that each spectral slice captured multiple cycles of the slowest rhythms yet still afforded fine temporal tracking of state transitions. Finally, the resulting power spectrogram (power vs. time and frequency) was aligned with our concurrently scored hypnogram, segmenting the 0.01–0.5 Hz power into Wake, NREM, and REM epochs, to directly compare how CSF oscillatory amplitude and bandwidth vary across sleep states. Within this spectral range, frequency ranges were defined as respiratory signal (RS, ~0.2–0.4 Hz), slow oscillation–linked NIRS oscillation (SONO, 0.6–0.7 Hz), slow oscillation–linked NIRS oscillation (SONO, 0.6–0.7 Hz), and cardiac cycle (CC, ~0.8–1.2 Hz). For Fig. 5A, Δ[H₂O] was band-pass filtered within each range and Hilbert-based instantaneous power was computed over time. For PSD-based peak analyses (Figs. 4N and 5, B to D), power spectral density was estimated from the Δ[H₂O] signal using Welch’s method with a 512-sample Hamming window, 50% overlap, and an FFT length of 1024; spectral peak power was defined as the maximum PSD value within each predefined band.

Human subject study

Multiple healthy subjects (N = 4, all male, age range: 28–37 years) participated in the study. The participants had no history of diseases that could affect the experimental results. Soft wearable NIRS and EEG/EOG measurements were acquired simultaneously from the same participants, with a total of 16 overnight recordings collected across all subjects, while independent datasets were used for sleep-stage model training. The experimental protocol (#H20211) was approved by the Georgia Tech Institutional Review Board, ensuring compliance with ethical research standards. In accordance with ethical guidelines, all participants provided written informed consent before the study. They were given a detailed explanation of the experimental protocol, potential risks, and their right to withdraw at any time.

Supplementary Materials

The PDF file includes:

Figs. S1 to S8

Tables S1 to S8

Legend for movie S1

Other Supplementary Material for this manuscript includes the following:

Movie S1