Respiration and heart rate from Wi-Fi CSI: what works, what does not

2026-10-11 · 2 min read

The idea is simple: chest movement changes multipath, and that shows up in CSI amplitude/phase (see the CSI sensing survey ^[1]). Engineering it is hard, because you must extract 0.1–0.5 Hz (breathing) and 0.8–2.0 Hz (heart rate) from noisy data.

Sampling is usually not the bottleneck

Our measured ~95 records/s across 8 chains is far above the Nyquist needs of a 2 Hz signal. The real limits are:

1. phase noise (CFO/SFO, hardware drift),
2. body motion (talking, shifting) swamping micro-motion,
3. too-short observation windows (spectral resolution).

Runnable pipeline

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

def bandpass(x, fs, lo, hi):
    n = len(x); X = rfft(x - x.mean()); f = rfftfreq(n, 1 / fs)
    X[(f < lo) | (f > hi)] = 0
    return np.fft.irfft(X, n)

def estimate(signal, fs=95.0, band=(0.1, 0.5)):
    n = len(signal); x = bandpass(signal, fs, *band)
    spec = np.abs(rfft(x * np.hanning(n))); freqs = rfftfreq(n, 1 / fs)
    mask = (freqs >= band[0]) & (freqs <= band[1])
    return freqs[mask][np.argmax(spec[mask])] * 60.0   # bpm

Order matters: unwrap → remove linear CFO/SFO drift → band-pass → window → peak — the canonical phase-based vital-signs work is PhaseBeat ^[3]. Resample by timestamp; never assume a constant frame rate.

When it can work

Honest limits

1. Not a medical device — no diagnostics, no medical claims;
2. heart rate is much harder than breathing (smaller amplitude, harmonics interference);
3. parameters must be re-calibrated per room/pose;
4. our platform's vitals module is gated (returns Unavailable when motion/SNR conditions fail) — that is more honest than printing a number anyway.

Published work reports the same distance/environment dependence: FarSense extended the respiration range using the CSI ratio of two antennas ^[2] — "it can sense" and "it can sense in any room" are different claims.

Related: Python pipeline · datasets · Wi-Fi CSI explained