Wi-Fi respiratory monitoring and heart rate from CSI: what works

By Wu Jun · Updated 2026-11-02 · 8 min read

Chest movement and heartbeats can change multipath, and that modulation may appear in CSI amplitude or phase ^[1]. A peak in a spectrum is not proof of a physiological signal, however. Fans, HVAC, walking, chain switching and hardware phase drift can all create a plausible-looking number.

This page gives a reproducible engineering route: calculate the true independent measurement rate, choose a relatively stable signal, resample by timestamp, use 30-second windows, gate unusable outputs and calibrate against an external reference. It is not medical advice, and the open dataset is not evidence of clinical accuracy.

0. Bottom line

The defensible design is not "print a bpm for every window." It is to emit a value together with quality metadata, or return Unavailable. Accuracy must be measured against a synchronised chest band, impedance device, pulse oximeter or manual count.

1. Record rate is not independent measurement rate

Our 10-minute capture contains 57,293 chain records, or about 95.5 records/s. One CSI measurement normally contains four or eight chain records that share the same sampling instant. Splitting on chain_info bit 15 gives 7,609 complete measurement groups, so the independent measurement rate is:

7,609 / 599.924 s ≈ 12.68 measurements/s

The full analysis script also removes 25 zero-heavy groups, using 7,583 measurements and obtaining about 12.64 Hz. This is above the 4 Hz Nyquist minimum for a 2 Hz heart-rate component, but the margin is not enormous. Spectral resolution, packet loss, timestamp jitter and window length still determine whether a result is usable.

Before plotting a spectrum:

1. close measurements on chain_info bit 15;
2. select one valid chain per group, or construct a two-antenna ratio;
3. sort by the actual timestamp instead of inventing time from a record index and a fixed frame rate;
4. inspect missing chains, 64-subcarrier records, all-zero I/Q and gaps longer than one second.

2. Which signal should you use?

Amplitude is easy to consume, but automatic gain, rate control and gross motion affect it. Phase contains finer displacement information, yet commodity NIC phase includes carrier and sampling frequency offsets plus hardware drift. Unwrapping one chain does not automatically remove those errors.

PhaseBeat uses CSI phase for respiration and heart-rate monitoring ^[2]. FarSense uses the CSI ratio of two antennas to suppress common-mode noise and extend respiration range ^[3]. A practical order to test is:

1. phase difference or complex ratio between two antennas in the same measurement;
2. phase difference between adjacent subcarriers;
3. first principal component of multi-subcarrier amplitude changes;
4. a single-chain amplitude baseline.

For every candidate, collect three classes first: empty room, still person and moving person. Check whether the feature tracks respiratory cycles rather than only detecting motion.

3. A deliberately conservative estimator

The functions below assume the input is a real-valued series produced by selecting one chain for each measurement. The estimator removes a linear trend, applies a Hann window and reports peak frequency plus two quality markers. Resample on the real timestamp grid first:

import numpy as np
from numpy.fft import rfft, rfftfreq

def resample_by_time(x, t, fs=20.0):
    x = np.asarray(x, dtype=float)
    t = np.asarray(t, dtype=float)
    order = np.argsort(t)
    t, x = t[order], x[order]
    tu = np.arange(t[0], t[-1], 1.0 / fs)
    xu = np.interp(tu, t, x)
    return xu, tu

def estimate_bpm(x, fs, band):
    x = np.asarray(x, dtype=float)
    n = len(x)
    if n < int(fs * 20):
        return None

    idx = np.arange(n, dtype=float)
    slope, intercept = np.polyfit(idx, x, 1)
    y = x - (slope * idx + intercept)
    y = (y - y.mean()) * np.hanning(n)

    power = np.abs(rfft(y)) ** 2
    freqs = rfftfreq(n, 1.0 / fs)
    mask = (freqs >= band[0]) & (freqs <= band[1])
    if not mask.any():
        return None

    band_power = power[mask]
    band_freqs = freqs[mask]
    k = int(np.argmax(band_power))
    peak_freq = band_freqs[k]

    floor = float(np.median(band_power)) + 1e-12
    prominence_db = 10.0 * np.log10((band_power[k] + 1e-12) / floor)
    band_ratio = float(band_power.sum() / (power.sum() + 1e-12))

    return {
        "bpm": float(peak_freq * 60.0),
        "prominence_db": float(prominence_db),
        "band_power_ratio": band_ratio,
    }

Example bands:

# Interpolation regularises time; it does not add physiological information.
xu, tu = resample_by_time(signal, timestamps, fs=20.0)

breath = estimate_bpm(xu, fs=20.0, band=(0.10, 0.50))  # 6–30 bpm
heart  = estimate_bpm(xu, fs=20.0, band=(0.80, 2.00))  # 48–120 bpm

A 30-second window has a theoretical bin spacing of 1 / 30 = 0.033 Hz, or about 2 bpm. That is not a promise of 2 bpm error. SNR, window leakage and non-stationarity still matter.

4. Thirty-second windows and availability gating

Start with a 30-second window and a 5-second hop. Respiration is covered by at least three cycles, while heart rate contains more. The tradeoff is at least 30 seconds of warm-up. Store quality fields with every estimate:

window = {
    "start_s": start,
    "end_s": end,
    "valid_sample_ratio": valid_count / expected_count,
    "max_gap_s": max_gap,
    "motion_score": motion_score,
    "band_power_ratio": result["band_power_ratio"],
    "prominence_db": result["prominence_db"],
    "bpm": result["bpm"],
}

An example gate:

def vitals_status(w, limits):
    if w["valid_sample_ratio"] < limits["min_valid_ratio"]:
        return "Unavailable: data coverage"
    if w["max_gap_s"] > limits["max_gap_s"]:
        return "Unavailable: gap"
    if w["motion_score"] > limits["max_motion"]:
        return "Unavailable: motion"
    if w["prominence_db"] < limits["min_prominence_db"]:
        return "Unavailable: weak peak"
    if w["band_power_ratio"] < limits["min_band_ratio"]:
        return "Unavailable: broadband noise"
    return "OK"

Do not copy these thresholds as universal constants. Collect labelled empty-room, still, talking, walking and posture-change data on the actual firmware, channel, link and installation. Choose thresholds on that data and report both false-positive rate and coverage.

5. A reproducible calibration protocol

This is an experimental design, not a completed accuracy result. Participants must give informed consent and be told that the work is for research, not diagnosis.

5.1 Setup

• fix router channel, bandwidth, transmit power and antenna orientation;
• record distance, subject orientation, wall materials and furniture layout;
• use a dedicated AP interface to produce steady traffic;
• place a reference device where its clock can be synchronised: chest band, pulse oximeter or an independent breathing-count application;
• record two minutes of an empty room to find fan, HVAC and power-supply cycles.

5.2 Capture sequence

1. empty room for two minutes;
2. seated and still for two minutes, without talking or changing posture;
3. paced breathing at 10, 15 and 20 breaths/min, at least one minute per segment;
4. normal breathing for two minutes with a synchronised reference;
5. small in-place motion for 30 seconds, followed by 60 seconds still;
6. repeat after changing distance or orientation, at least three times.

Do not use "I feel like my heart rate is 70" as ground truth. Use a time-synchronised pulse oximeter or another validated reference for research comparison only.

5.3 Store with every run

6. How to report metrics

Reporting only the window with the smallest error against a reference overstates performance. At minimum report:

• coverage: fraction of all windows that pass the quality gate;
• conditional MAE/RMSE: computed only on OK windows with a valid reference;
• empty-room false-positive rate: how often a normal rate appears with nobody present;
• recovery time: delay before a stable estimate returns after motion;
• latency: window length plus hop delay;
• results by scenario: distance, orientation, subject and device separately.

When comparing two algorithms, use the same windows, reference timeline and gate. If one method reports only 20% of windows and has a lower MAE, show both the error and the coverage.

7. Common failure modes

8. What the public dataset can and cannot prove

The public 10-minute capture is useful for:

• parsing the 24-byte CSI header and I/Q;
• grouping records on bit 15;
• building amplitude heatmaps and motion metrics;
• testing timestamps, loss and chain-selection code.

It has no synchronised chest-band or heart-rate reference, so it cannot support claims such as "2 bpm respiration error" or "5 bpm heart-rate error." Those results require a new consented capture with a reference device.

9. Minimum implementation checklist

• select one chain per measurement; do not treat 95.5 records/s as a 95.5 Hz physiological signal;
• resample on timestamps and record the maximum gap;
• establish empty-room and still-person baselines before trying respiration;
• repeat each subject, posture and distance at least three times;
• keep raw CSI, reference values, labels and quality metrics;
• prove respiration before attempting heart rate;
• return Unavailable when the gate fails;
• make no diagnosis, disease-monitoring or guaranteed-accuracy claim.

10. Related implementation

• Phase processing and spectral estimation: this page plus the Python pipeline;
• Real data: 10-second sample and 10-minute dataset;
• Dataset parsing: from eight parts to a heatmap;
• Published methods: PhaseBeat ^[2], FarSense ^[3] and the CSI sensing survey ^[1].