World Sleep Industry ENCYCLOPEDIA

UWB radar respiration and heartbeat detection algorithms

4.4.1

Definition and classification boundary

UWB radar respiration and heartbeat detection algorithms use the range resolution of ultra-wideband radar to track minute displacement at the chest, abdomen, neck, or another body surface caused by breathing and cardiac mechanical activity. Outputs may include respiratory rate, heart rate, candidate breathing pauses, beat intervals, and signal-quality measures. A common implementation uses impulse-radio UWB and observes changes over time within the same range gate across successive radar frames. It requires no skin electrode or respiratory belt, but measures mechanical motion rather than ventilation volume, airflow, oxygen saturation, or cardiac electrical activity itself.

UWB describes a signal occupying a broad spectrum or having a large fractional bandwidth, while millimeter wave describes carrier frequency; the terms are not mutually exclusive opposites. This entry covers the specialized recovery of respiration and heartbeat from UWB returns. The broader millimeter-wave sleep-monitoring stack, sleep staging, dedicated respiratory-event sensor modules, and generic presence detection have separate boundaries. Whether a product has a medical function depends on its intended use and how its output affects decisions, not merely on the labels “contactless” or “ultra-wideband.”

Range profiles, target lock, and preprocessing

An impulse-UWB system can arrange the echoes following each transmitted pulse by propagation delay to form a range profile. Repeated frames form a fast-time by slow-time matrix: fast time maps to range, while slow time captures temporal change. Processing commonly includes antenna-delay and baseline calibration, static-background removal, abnormal-frame rejection, range-gate selection, target tracking, and resampling. Walls, bed frames, and furniture create fixed and multipath returns, so the system must distinguish the direct human reflection from longer-path mirror components.

Target locking cannot be a startup-only operation. A turn, sitting up, leaving the bed, arrival of a bed partner, or movement of bedding can shift the strongest gate. The tracker should reacquire the person using spatial continuity, motion pattern, and physiological micromotion together. Every interval should retain target range, echo energy, interference level, identity continuity, and validity state. If people cannot be separated or tracks exchange identity, the algorithm should reject the estimate or mark it unknown rather than splice the respiration and heartbeat of two people into one record.

Respiratory extraction and event candidates

Breathing produces relatively large, slow, periodic thoracoabdominal displacement. It can be estimated from a background-corrected range-gate signal using band-pass filtering, phase or amplitude demodulation, autocorrelation, spectral-peak tracking, and time–frequency analysis. The algorithm must also handle baseline drift, posture-dependent amplitude, sighing, speech, coughs, and irregular breathing. A short sliding window has poor frequency resolution, whereas a long window delays its response to change; an output should therefore carry its window duration, update rate, and uncertainty.

A sudden reduction in respiratory amplitude may reflect a genuine event, but it may also result from lateral obstruction, a body turn, loss of the range gate, or recovery after gross motion. Chest-wall mechanics alone generally cannot establish whether airflow stopped, oxygen fell, or an EEG arousal occurred, and may not robustly distinguish obstructive from central events. Event candidates should preserve coverage and motion flags. When used for apnea screening, the complete device requires validation in the intended population; a single period with no detected displacement must not be translated directly into “no breathing.”

Heartbeat extraction and cardiorespiratory separation

Cardiac displacement is usually much weaker than respiration and is readily obscured by the respiratory fundamental and harmonics, bed-frame vibration, and subtle posture change. Processing within the target gate may use band decomposition, harmonic suppression, cepstral analysis or autocorrelation, time–frequency tracking, blind-source separation, and probabilistic continuity constraints, with adjacent gates compared to select a stable component. The algorithm must avoid labeling a second or third respiratory harmonic as heart rate and recognize cases in which rapid breathing and a low heart rate bring the bands close together.

Average heart rate, beat-to-beat intervals, and rhythm classification are distinct tasks. A dominant window frequency that agrees with ECG heart rate does not establish accurate localization of every beat and is insufficient for reliable heart-rate variability or diagnosis of arrhythmias such as atrial fibrillation. Beat-level validation should measure missed and duplicate beats, timing error, and valid coverage. After gross movement, respiration may safely resume before heart rate; interpolating the gap merely manufactures a continuous interval series.

Evidence and validation design

Studies using synchronized polysomnography support the feasibility of IR-UWB respiratory and heart-rate estimation under particular conditions, but sample size, posture, population, and algorithms remain limited. One study compared radar with respiratory belts and ECG in six healthy volunteers and 15 people with suspected sleep apnea. It found promising agreement within groups, while some heart-rate agreement declined with increasing disease severity. Another 176-participant study reported strong association between radar-derived and PSG-derived AHI, yet night-level correlation or classification cannot replace event-level error analysis, external home validation, or evidence of clinical outcomes.

Respiratory rate should be compared with synchronized airflow, effort belts, or carbon-dioxide references. Heart rate and beat intervals should be compared with ECG. Reports should include mean absolute error, error distributions, Bland–Altman limits of agreement, valid coverage, and recovery time after motion. Event tasks additionally require sensitivity, specificity, precision, recall, false alarms per night, and duration error. Splits must isolate participants and disclose whether invalid intervals remain in the denominator; evaluating only motionless, measurable windows overstates whole-night performance.

Deployment, safety, and data governance

Deployment specifications should define distance, angle, height, field of view, allowed obstruction, and recalibration conditions relative to the chest and abdomen, and should cover supine, lateral, and prone positions, different body sizes, and bed types. Pets, partners, fans, curtains, metal frames, nearby movement, and equipment vibration can all enter the return. The system should distinguish “target absent,” “insufficient signal,” “gross motion,” “multi-person conflict,” and “suspected physiological abnormality,” alert on prolonged unmeasurability, and permit confirmation by a user or caregiver.

Spectrum rules, emission limits, and device authorization vary by jurisdiction and intended use; hardware or firmware changes can also shift the return distribution and model performance. Diagnostic, monitoring, or treatment-triggering uses need a defined intended use, risk controls, version traceability, fail-safe degradation, and target-population validation. Although radar data are not ordinary images, they can reveal bedroom occupancy, routines, respiration, and cardiac activity. Products should minimize retention of raw returns, restrict debug interfaces, secure transmission and cloud accounts, and explain the sensing boundary to partners and care recipients.

Product acceptance and future directions

Acceptance criteria should be set separately for respiratory rate, heart rate, beat intervals, and respiratory events; one aggregate “accuracy” cannot represent them all. Procurement should also examine operating band and bandwidth, range resolution, frame rate, antenna arrangement, maximum target count, edge-compute capacity, missing-data encoding, raw-data access, and software update and rollback policy. Test sets should span complete nights and failure scenarios, with results stratified by participant, night, posture, signal quality, and disease severity.

Future directions include multi-antenna localization, durable multi-person identity tracking, physics-informed joint cardiorespiratory models, adaptive harmonic separation, calibration across rooms, and uncertainty estimation. Fusion with oximetry, ECG, mattress pressure, or audio can provide redundancy, but it also requires time synchronization, person matching, and additional privacy controls. A trustworthy system treats unmeasurable intervals as a formal output and tells users when a trend is informative, when installation needs correction, and when a standard test should take over.