Monitoring Breathing via Signal Strength in Wireless Networks

Introduction

• Potential uses include monitoring post-surgical respiratory risk,,,, sleep apnea, and SIDS, as well as rescue or occupancy detection; sensing motionless people also raises privacy concerns,.The paper notes that parents of infants at high risk for SIDS may want an alert if breathing becomes depressed or stops.Source: §1 · pp. 1–2
• In an otherwise stationary environment, the collective data from static wireless links can reliably reveal breathing, despite small breathing-related changes in RSS being difficult to detect on an individual link.The authors observed RSS changes when a person stood on or near a link’s line of sight, and a network-wide spectral peak close to the measured breathing rate.Source: §1 · p. 1
• BreathTaking detects a strong RSS frequency in the human breathing range rather than measuring exhaled gases; other cyclic motion in that range could produce a similar signal, so context matters.The considered range is 10–40 bpm (0.167–0.667 Hz).Source: §1 · pp. 1–2
• With 30 seconds of RSS data from a 20-device network, rate estimates had 0.1–0.4 bpm RMSE, and directional-antenna experiments had no false alarms or missed detections.The experiments monitored an otherwise motionless person in a hospital room with no other person present.Source: §1 · p. 2
• The method estimates breathing rate with maximum likelihood under a sinusoidal-signal approximation, using RSS measurements from many links simultaneously to detect breathing.The paper also examines rate, amplitude, and phase estimation and evaluates performance against network size, person position, and breathing rate.Source: §1 · p. 2

2 METHODS

• The methods define network RSS measurements and model breathing through a periodic component in link signals.The component is assumed to have the same frequency across links, while its amplitude and phase can differ.Source: §2 · p. 2, §2.2 · p. 2

2.1 Network

• Sampling above 4/3 Hz is required under the assumed maximum breathing rate of 40 breaths per minute.The paper identifies this as the Nyquist sampling rate for that assumed maximum.Source: §2.1 · p. 2
• Each measured link l is an ordered transmitter–receiver pair, and its RSS at time i is denoted yl(i).Source: §2.1 · p. 2
• The network need not be fully connected; the experiments use a mesh topology to maximize the number of measured links, without excluding other topologies.Source: §2.1 · p. 2

2.2 Signal Model

• Without environmental motion, the signal model represents each link’s mean RSS with additive noise assumed to be zero-mean Gaussian, identically distributed over time, and independent across links.Source: §2.2 · p. 2
• With a breathing person present, each link is assumed to gain a sinusoidal RSS component with shared frequency across links and link-specific amplitude and phase.The sampling period is also assumed identical across links.Source: §2.2 · p. 2

2.3 Framework

• The observation period T corresponds to the RSS measurement duration before breathing detection and rate estimates are produced.The measurements are sampled with period Ts, and the observation period is tied to monitoring latency.Source: §2.3 · p. 3
• The framework’s objective is to detect whether a person is breathing within the network and estimate the model parameters θ from measurements across links.Source: §2.3 · p. 3
• Breathing frequency f is the parameter of greatest interest, and the modeled range is 0.167–0.667 Hz.Source: §2.3 · p. 3

2.4 DC Removal Filtering

• A 7th-order Chebychev high-pass filter with 0.1 dB maximum passband ripple and fmin = 0.167 Hz is applied to every link’s signal.Order 7 provides at least 40 dB attenuation below 0.1 Hz; the filtered signals are assumed to have zero mean.Source: §2.4 · p. 3
• The filter attenuates DC because it can hide lower-amplitude sinusoidal power, while a nearly flat gain above fmin avoids biasing frequency estimates.Breathing is assumed absent below fmin, and linear phase is not required.Source: §2.4 · p. 3

2.5 Breathing Estimation

• The MLE extends sinusoid parameter estimation to L link signals with separate amplitudes and phases, while constraining frequency to [fmin, fmax] [10, p. 193–195].The section introduces this estimation before discussing breathing detection.Source: §2.5 · p. 3
• Under the i.i.d. Gaussian noise model, maximizing likelihood is equivalent to minimizing J, and a good approximation to the frequency MLE applies when normalized frequency is away from 0 and 1/2 [10, p. 193–195].Frequencies close to zero are excluded, and sampling is significantly faster than the Nyquist rate.Source: §2.5 · p. 3
• The link amplitude and phase estimates are computed using the estimated frequency; frequency estimation from a single link is possible by setting L = 1.Source: §2.5 · p. 3

2.6.1 Single-link breathing detection

• Single-link detection uses one link’s RSS and assumes its estimated breathing amplitude is higher when a person is present, testing against a user-defined threshold.The frequency and amplitude estimates are calculated using the single-link case, L = 1.Source: §2.6.1 · p. 4

2.6.2 Network-wide breathing detection

• The network-wide statistic sums squared amplitude estimates across links and scales their average by the sample count to support a constant threshold under H0.The method uses a user-defined detection threshold; under H0, average squared link magnitude is approximately inversely proportional to the number of samples.Source: §2.6.2 · p. 4

2.7 Performance Analysis

• Detection performance is evaluated using false-alarm probability, the fraction of no-breathing experiments decided as H1, and missed-detection probability, the fraction of breathing experiments decided as H0.Both probabilities are desired to be zero, but there is a tradeoff between them.Source: §2.7 · p. 4
• Post-surgical monitoring prioritizes low missed detection, while search and rescue may prioritize low false alarms to avoid fruitless searches.The section explains that applications have different requirements for the two error probabilities.Source: §2.7 · p. 4
• Breathing-rate estimation is evaluated by the RMSE of the estimated frequency across K experimental realizations, comparing each estimate with its actual breathing frequency.For realization k, the estimate is ˆf(k) and the actual frequency is f(k).Source: §2.7 · p. 4

3.1 Network

• Twenty TelosB sensors measured each link’s RSS at 4.16 Hz, above the 0.667 Hz maximum breathing rate; packet data was collected by a hallway base station.Nodes transmitted sequentially using SPIN, and the base station time-stamped the data recorded on a laptop.Source: §3.1 · p. 4
• Experiments compared an omnidirectional 2.25 dBi dipole antenna with an 8.0 dBi directional patch antenna.Source: §3.1 · p. 4

3.2 Environment and Setup

• In the clinical-room setup, sensors were placed around the bed without connecting them to it, so breathing-related RSS periodicity would not result from sensor movement.The bed itself automatically changed pressure every 3–4 minutes, so a person lying in it was not perfectly stationary; sensors were positioned on tables, carts, and PVC-pipe stands beside the bed.Source: §3.2 · pp. 4–5

3.3 Breathing Rate

• Ground truth was checked by having participants follow a metronome rate and verifying it afterward against video of an end-tidal CO2 monitor’s displayed estimate.Participants breathed through their noses while connected to the monitor, which estimated breathing frequency from the CO2 signal.Source: §3.3 · p. 5

4.1 Single Link Breathing Monitoring

• At a zero-false-alarm threshold, single-link detection reached only 0.7%; of 63 high-amplitude estimates, just 17 were within 1 bpm of the true 15 bpm rate.The other 46 frequency estimates were nearly uniformly distributed over [fmin, fmax], so individual-link amplitude and frequency estimates are unreliable for monitoring.Source: §4.1 · p. 5

4.2 BreathTaking Rate Estimation

• For T ≥30 s, fewer than 2% of rate estimates were invalid; valid-estimate RMSE was below 0.5 bpm for T ≥25 s, bias was about 0.1 bpm, and no estimates were invalid for T ≥50 s.The experiments used patch antennas and controlled breathing at 15 bpm.Source: §4.2 · pp. 5–6, Figure 3 · p. 6

4.3 BreathTaking Amplitude Estimation

• Only 5% of links had ˆAl > 0.331, and high-amplitude links crossed the chest and had average RSS of -48.6 versus -39.9 overall.Their paths were only 13% longer on average, suggesting that links in a deep fade are more useful for breathing monitoring; adaptive use of deep fades is proposed as future work.Source: §4.3 · p. 6

4.4 BreathTaking Detection Performance

• With patch antennas, γnet = 1.50 perfectly distinguished the person-present and empty-room cases in the experiment, whose measured ˆS ranges did not overlap.The empty-room values ranged from 0.98 to 1.45; person-present minima were 1.57 at T = 15 s, 1.63 at T = 30 s, and 2.03 at T = 60 s.Source: §4.4 · p. 6, Figure 5 · p. 7

4.5 Antennas

• With hallway motion, patch antennas’ maximum empty-room ˆS rose to 1.52, while dipole antennas’ rose to 2.08 and overlapped person-present values at T = 30 s.Breathing remained monitorable with dipoles, but directional patch antennas were more robust to motion outside the room.Source: §4.5 · p. 7, Figure 7 · p. 7

4.6 Rate Changes; 4.7 Bed Height

• At T = 30 s, rate estimation performed best at 12 bpm without monotonically degrading at higher rates; all tested bed heights had RMSE below 0.5 bpm, with near-LOS about half the next-best RMSE.Bed-height tests used a 15 bpm breathing rate; every height had a very small chance of an invalid estimate.Source: §4.6 · p. 8, Table 1 · p. 8, §4.7 · p. 8

4.8 Fewer Sensors

• With 13 sensors, invalid rate estimates fell to 1.3% and valid-estimate RMSE was below 0.3 bpm; 19 sensors yielded zero invalid estimates, while seven performed poorly.The seven-sensor network had almost one in four estimates invalid; results averaged 100 randomly selected subsets for each network size.Source: §4.8 · p. 8, Figure 8 · p. 8

4.9 Phase Estimation

• Across five experiments, phases of the highest-amplitude links formed two modes separated by 180°.The paper suggests phase modes may help distinguish multiple stationary breathers or improve monitoring, but leaves those methods for future work.Source: §4.9 · p. 9, Figure 9 · p. 9

5 RELATED WORK

• Capnography measures exhaled carbon dioxide through tubes and estimates breathing rate from its signal frequency, whereas BreathTaking indirectly senses periodic body-size changes without physical attachment.Capnography is standard for anesthetized patients in emergency departments and intensive care units; detached tubes can trigger an apnea alarm, and its mask or nasal cannula may limit movement.Source: §5 · p. 9
• Plethysmography uses body-attached electrodes to measure inhalation- and exhalation-related changes in inductance or impedance, while BreathTaking needs neither body attachment nor wires.Plethysmography is used in home monitors for infants at risk of SIDS.Source: §5 · p. 9
• At sedation-procedure offices, pulse oximeters are used instead of capnography or plethysmography because of equipment costs, but detect oxygen desaturation only minutes after breathing ceases.Source: §5 · p. 9
• Unlike proposed microwave Doppler and UWB radar systems, BreathTaking uses lower-cost off-the-shelf wireless devices in a network of simple transceivers.UWB radar has also been proposed for unobtrusive vital-sign monitoring and may detect a stationary person’s heart rate.Source: §5 · p. 9

• BreathTaking monitors a stationary person non-invasively using RSS measurements across multiple wireless links, with a maximum likelihood estimator of breathing parameters and detection algorithms.The estimator uses all measured links’ RSS data simultaneously.Source: §6 · p. 9
• Using 30 seconds of measurements, BreathTaking estimates breathing rate with 0.1 to 0.4 bpm error.Source: §6 · p. 9
• Breathing detection distinguished breathing from its absence without false alarms or missed detections using 15 seconds with patch antennas and 30 seconds with dipole antennas.Source: §6 · p. 10
• Links with low average RSS were most affected by breathing, affected links’ estimated phases formed two modes separated by 180 degrees, and antenna directionality improved robustness to external motion.Source: §6 · p. 10
• Medical use requires extensive evaluation across many people and settings; further work could test other physical-layer protocols and center frequencies, and combine BreathTaking with other sensors and context information.Source: §6 · p. 10