A proposed system that reads breathing from commodity Wi-Fi reported 97% average accuracy in normal-breathing tests. Its mean absolute error, or MAE, a measure of the average size of the rate estimate's error, was 0.48 breaths per minute. Compared with the second-best algorithm, the reported gap was about 2.8 percentage points in accuracy and 0.45 breaths per minute in MAE.
The figures came from four volunteers in 58 author-collected tests: 28 normal-breathing tests and 30 breath-holding tests. A public dataset added 30 tests. The setup recorded channel state information, or CSI, from 256 subcarriers on 5 GHz Wi-Fi with 80 MHz bandwidth at 50 Hz, using a Neulog respiration belt as the reference.
For the headline comparisons, the paper did not report confidence intervals or dispersion estimates. The figures therefore describe the tested comparisons without a reported range showing how much performance might vary.
A second dataset showed a similar result
On the public dataset, the proposed method recorded 97% accuracy and a 0.4 breaths-per-minute MAE. RMS was reported at 95% accuracy and 0.6 breaths per minute. WiRe and FFT were each around 92% accuracy and around 1 breath per minute in MAE, while Savitzky-Golay filtering, or S-G, recorded 88% accuracy and 1.6 breaths per minute in MAE.
How the signal was handled
The proposed framework models different sensitivities among CSI subcarriers, selects informative subcarriers, estimates breathing rate and extends the approach to breathing-pause scenarios.
Its selection module constructs a four-dimensional feature vector and applies K-means clustering with K set to 3. It selects the cluster with the highest average energy in the breathing band, then chooses the sensitive subcarrier with the highest fit-quality value.
Amplitude and phase are handled separately before the estimates are fused. The system retains the candidate with the higher fit quality, then uses a three-second sliding window and an amplitude threshold of eta = 0.3 to detect a pause. The threshold is expressed as a proportion of the reference amplitude.
Pause episodes were a separate test
In the pause tests, the false-detection rate, meaning pauses called when none was present, and the missed-detection rate, meaning pauses the system failed to catch, were both below 10% at eta = 0.3. The reported threshold sweep put each at about 8%. At eta = 0.5, the duration error was around 0.5 seconds, roughly 18% of the three-second analysis window.
During detected pause episodes, the method reported 88% breathing-rate accuracy and a 1.6 breaths-per-minute MAE. WiRe reported 78% accuracy and 3 breaths per minute in MAE, while the other comparison methods had MAEs above 5 breaths per minute. The paper described the performance difference over baselines as more than 9%.
A separate check used eta = 0.3 unchanged in 10 additional tests involving two new volunteers. The transceiver distance was 1.5 metres; false detections were 7% and missed detections were 4%, both below 10%. The added check was small and scenario-specific.
The analysis notes that the threshold was derived from the current dataset and may require recalibration across deployment environments.
Results across posture and distance
Across four lying postures, right-side, left-side, supine and prone, the proposed method recorded 96% average accuracy and 0.63 breaths-per-minute average MAE. Versus WiRe, the reported differences were 1.6 percentage points in accuracy and 0.24 breaths per minute in MAE.
At the two tested distances, the proposed method's accuracy was above 90% within 10 metres and was 84% at 20 metres. The all-method average was around 80% at 10 metres and below 60% at 20 metres; WiRe recorded 65% at 20 metres.
In ablation comparisons, reported accuracy was 1.3 points lower without normalization, 2.4 points lower when sinusoidal fitting was replaced with FFT, and 15 points lower with arbitrary subcarrier selection. The amplitude-and-phase fusion variant was 2 points higher than amplitude-only processing and 7 points higher than phase-only processing.
The reported implementation processed one minute of CSI in an average of 3.26 seconds, with packet processing taking about 1 millisecond and memory use below 100 megabytes. The paper gives its complexity as O(N · K · G) and describes it as linear in input size.
The limits of the evidence
The study is an engineering evaluation, not a clinical validation. Its evidence covers four volunteers, 58 author-collected tests and a 30-test public dataset, with normal breathing, voluntary breath-holding, selected lying postures and tested distances. It does not show how the system would perform for other people, irregular breathing patterns, multiple people or highly dynamic environments.
The pause analysis also assumes abrupt transitions between breathing and silence and a single dominant breathing frequency. The additional threshold check involved only 10 tests, so the reported error rates should be read as results from this evaluation rather than a general guarantee across deployments.
The supplied document is an arXiv preprint, version 1, dated 26 August 2026. It reports partial support from the Smart Traffic Fund established under the Transport Department of the Hong Kong Special Administrative Region, China.
The results point to an engineering approach that warrants broader testing for contact-free respiratory monitoring. They do not establish clinical readiness or apnea diagnosis.
Paper data and sources
Original title: A Subcarrier-Aware Approach for Robust Respiratory Monitoring with Commodity Wi-Fi
Authors: Pei Tang, Yunpeng Ge, Ivan Wang-Hei Ho
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text