Name, Definition, and Category Boundary
A Wi-Fi CSI sleep respiration and body-movement monitoring device consists of one or more wireless transmitters, receivers, and a processing system. It analyzes changes over time in Wi-Fi channel state information within a sleep area to infer respiration-related micro-motion, larger body movements, and selected derived states without requiring a worn sensor. CSI describes the complex response of a wireless channel for specified transmit antennas, receive antennas, frequencies, or subcarriers. Chest and abdominal motion, turns, and environmental changes alter reflection, scattering, blockage, and multipath, leaving variations that may be analyzed.
The category does not mean that ordinary router software alone provides reliable monitoring. An implementation may use a dedicated pair of nodes, a commercial network interface that exposes CSI, a development platform with modified firmware, or newer hardware supporting standardized WLAN sensing procedures. It also requires defined packet scheduling, clock behavior, calibration, algorithms, and a user interface. A system that observes only RSSI, link rate, or packet-loss statistics should not automatically be described as CSI-based, while passive Wi-Fi radar uses a different reference-channel and signal-processing architecture even if the illuminating waveform is Wi-Fi.
This entry is limited to devices that explicitly use Wi-Fi CSI to estimate sleep respiration and body movement. Entry 200, wireless channel-sensing sleep monitoring devices, is broader and can include other radio technologies, channel measures, and sensing methods. Wi-Fi CSI sensing is also distinct from millimeter-wave radar, cameras, acoustics, mattress pressure sensors, wearables, and polysomnography. Having no screen, requiring no worn device, or reusing a communications network does not remove responsibilities for installation, validation, privacy, and clear intended-use labeling.
CSI Principles and Signal Meaning
In a Wi-Fi link using OFDM and multiple antennas, a receiver can estimate a complex channel response across transmit antennas, receive antennas, and subcarriers; amplitude and phase together characterize propagation. CSI is finer grained than RSSI, which compresses reception across a channel into a coarse value, but CSI remains a link measurement rather than an anatomical image. Any moving person, bedding, curtain, fan, door, or piece of furniture may change multipath. An algorithm therefore models a physical or statistical relationship between propagation changes and a target state; it does not observe that state directly.
Millimeter-scale chest and abdominal motion during quiet breathing can periodically alter selected paths, while turning, sitting, and limb motion cause larger and broader-band changes. Visibility depends on whether the target lies in a geometrically sensitive region, how reflected paths combine in phase, and antenna polarization and direction. Being closer to a node does not guarantee greater sensitivity, and some locations create low-sensitivity zones or blind spots. The same respiratory movement can produce different amplitude and phase trajectories at another bed position, posture, room, frequency band, or antenna link.
CSI does not directly measure nasal or oral airflow, thoracoabdominal effort, oxygen saturation, ECG, EEG, or sleep stage. A respiratory rate derived from a periodic component estimates mechanical respiratory motion, while an apparent pause may instead result from turning, blockage, link interruption, or movement into a blind spot. A movement magnitude is not limb force, stillness is not sleep, and absence of CSI variation does not by itself establish an empty room, absent breathing, or cessation of vital signs. Every derived output should preserve signal-quality and failure-state information.
Hardware, Link, and Deployment
Deployment should specify transmitter and receiver models, firmware, antenna count and orientation, band, channel, bandwidth, transmission schedule, packet rate, CSI format, and timing source. Chipsets expose different subcarrier sets, quantization, scaling, and phase information, and a driver or firmware update can change fields or sampling behavior. Stating compatibility with a Wi-Fi generation, router brand, or standard is not a substitute for validating the exact hardware combination. Compliance with a WLAN sensing protocol supplies neither a medical algorithm nor evidence of sleep-monitoring accuracy.
Transmitter and receiver placement defines the sensing region. Instructions should state height, distance, orientation relative to the bed, line-of-sight or non-line-of-sight assumptions, and allowed wall and furniture changes, and installation should verify both link quality and target coverage. Moving a node to the other side of the bed, shifting furniture or the bed, closing a door, adding metal, or changing a bed partner's location can reorganize multipath. Automatic calibration can establish the current environmental baseline, but an unseen major layout change should not be silently absorbed while the system continues to report high confidence.
When communication and sensing share radio resources, the device must manage user traffic, sensing frames, roaming, power save, automatic channel selection, bandwidth changes, rate adaptation, and neighboring-network interference. Network load, access-point restart, client sleep, or dynamic frequency selection can make CSI sampling sparse, irregular, or unavailable. Documentation should explain whether sensing depends on continuous traffic, whether it affects network throughput, and which functions remain during internet loss, local-network operation, or total link loss. An interpolated uniform sequence must not be represented as continuously acquired measurements.
Data Processing and Multi-Person Failure
Raw CSI commonly contains carrier-frequency offset, sampling-frequency offset, packet-detection delay, automatic-gain effects, phase bias, missing packets, and outliers. Processing may include parsing and scaling, amplitude or phase calibration, detrending, time resampling, antenna or subcarrier ratios, principal-component or blind-source separation, spectral or time-frequency analysis, and signal-quality scoring. Preprocessing changes visible frequencies and phase, so it must remain tied to hardware, firmware, sampling behavior, and algorithm version. Tuning and evaluating on the same recordings exaggerates generalization.
Gross movement is usually much stronger than respiratory micro-motion. Respiratory estimation should pause or become invalid during movement and relock only after the channel stabilizes. Fan blades, robots, pets, curtains, doors, caregivers, and activity in an adjacent space can generate periodic or transient changes. A dominant frequency within a plausible respiratory range does not establish that the source is the person in bed. The algorithm should distinguish no target, a still target, motion contamination, low-sensitivity placement, wireless interruption, and unknown interference, and report initial lock time, recovery, valid coverage, and the longest gap.
Reflections from multiple people are superimposed at the receiver. When respiratory rates are similar, positions change, or people move together, spectral peak separation and blind-source separation can fail. Even two recovered periodic components may not be reliably mapped to a name or side of the bed. Multiple links, antenna arrays, and spatial zoning may improve identifiability but do not guarantee attribution. A device not validated for subject count, bed-side swaps, similar rates, and partner entry or exit should be limited to one person or explicitly return unattributable results.
Sleep Outputs and Clinical Validation
Foundational outputs may include respiratory rate, a respiration-related waveform, movement, turns, stillness, bed exit, or presence. Sleep onset, wake, total sleep time, sleep efficiency, stage, apnea, and a sleep score are additional inferences. Quiet wake and sleep may both show regular breathing with little movement, and brief arousals may provide too little wireless evidence. Without EEG, EOG, and chin EMG, the device does not directly observe standard staging signals. A high laboratory classification score does not make the result equivalent to polysomnography.
Respiratory rate should be validated against a synchronized respiratory reference appropriate to the setting, movement against independent annotation or another sensor, and sleep or respiratory-event claims against the relevant polysomnographic channels and clinical scoring. Time alignment must account for wireless timestamps, buffering, resampling, and algorithm windows. Evaluation should report interval-level error, sensitivity and specificity, confusion matrices, event definitions, valid coverage, missing data, false alarms, and recovery, rather than only controlled-breathing classification accuracy or aggregate correlation.
Testing should span people, body sizes, ages, postures, respiratory rates, nights, rooms, bed positions, furniture, node locations, hardware lots, wireless congestion, and single- and multi-person conditions, with training, tuning, and independent testing kept separate. A few healthy participants breathing on command or resting briefly establish feasibility only; they do not support abnormal-breathing, sleep-apnea, pediatric, older-adult, or clinical-alarm claims. A model can undergo domain shift in a new room or after a network interface is replaced, and on-site calibration does not replace independent external validation.
Operations, Safety, and Reliability
Commissioning should associate each node with the correct room and bed, verify coverage, run self-tests, synchronize clocks, and establish an empty-room baseline, with a visible installation-failure state. Runtime monitoring should cover transmitters, receivers, antennas, power, clocks, packet rate, channel, firmware, storage, and processing services. Automatic channel selection, router replacement, credential changes, factory reset, software update, or furniture rearrangement should trigger impact assessment and any needed recalibration or verification instead of silent reuse of an old model.
Reliability testing should cover long operation, temperature, power variation, network congestion, adjacent-channel interference, packet loss, restart, full storage, clock drift, displaced nodes, and remote-service outage. Coexistence testing should evaluate sensing and communications while Bluetooth, neighboring Wi-Fi networks, and co-channel devices operate. Radio transmission, electromagnetic compatibility, electrical safety, and human exposure must meet requirements applicable to the sales region, band, power, antenna, and installation distance. Integration of an authorized radio module does not automatically cover every antenna or finished-product change.
If outputs participate in an alarm, validation must also distinguish technical failure from physiological abnormality and cover thresholds, latency, repeated notification, silencing, acknowledgment, escalation, and local degradation. A link interruption or blind spot must not appear as breathing cessation, and absence of an alarm must not be marketed as proof of safety. Routine network maintenance should not terminate important monitoring without an obvious status indication; conversely, a product without medical-purpose evidence should not use interface colors, push-notification wording, or marketing to make a general-wellness trend resemble emergency monitoring.
Privacy, Cybersecurity, and Use Limits
Wi-Fi CSI produces no visible photograph, yet it can infer presence, motion, routine, and respiratory trends while a person wears nothing and may not actively use a client device. It is therefore not inherently free of privacy risk. The sensing area may extend beyond the bed and may respond to activity behind a wall or in an adjacent space. Deployment should disclose who may be sensed, when sensing operates, its spatial scope, purposes, and retention, and provide suitable permissions and opt-out handling for bed partners, visitors, caregivers, and affected neighbors.
The system should protect node and account identity, wireless access, management interfaces, raw CSI, derived indicators, transmission, storage, logs, models, and software updates. Encryption of communication content does not hide every physical-channel variation; forged motion, jamming, malicious nodes, or replayed data can also corrupt results. Least privilege, unique credentials, trusted updates, vulnerability response, deletion, and cybersecurity-state awareness should span the lifecycle, with responsibilities separated among local sensing, cloud analytics, and third-party network services.
Without evidence and authorization for the particular purpose, this device category cannot diagnose sleep apnea, distinguish obstructive from central events, measure oxygen saturation, identify arrhythmias, or replace polysomnography. It must not serve as the sole means of infant safety, SIDS prevention, life support, emergency response, or nurse call. Severe breathing difficulty, cyanosis, chest pain, fainting, unresponsiveness, or another danger sign requires action through the applicable emergency pathway; a person should not wait for wireless confirmation or use a normal trend to rule out disease.