How Many Antennas Does It Take to Stop Missing Atrial Fibrillation at Home?

Many heart-rhythm problems come and go. Paroxysmal atrial fibrillation (AF), an irregular rhythm that starts and stops on its own and raises stroke risk, is routinely missed because a clinic ECG samples ten seconds of a patient's life. Contactless radar placed by the bed promises round-the-clock coverage without electrodes, and recent systems detect AF from chest-wall motion alone. The catch is a quieter version of the same undersampling problem. Ambient sensors lose the signal whenever the patient turns away, moves, or sits in a spot where reflected waves cancel. Those windows are thrown out as "undecidable," and events that happen inside them go unreported. A clinical wearable study discarded 32.4% of all 30-second windows this way (Jeanningros et al.).

This proposal treats the fraction of time a patient is actually observable as the quantity to engineer. The idea comes from large-array wireless communication. When a receiver adds up signals from many antennas, the random fading seen by any single antenna averages out, so the combined signal strength becomes nearly constant. Engineers call this channel hardening. We ask whether the same averaging turns radar observability from a coin flip into a dependable quantity, and how fast observability improves as antennas are added. We also ask what hardening cannot fix: a torso turned away blocks every antenna in a single co-located array at once, so we expect a floor that only spatially distributed antennas can break.

Core Claim

For ambient cardiac monitoring, underreporting of paroxysmal events is driven by coverage at least as much as by how accurate the detector is on the windows it keeps. Coverage should scale predictably with spatial diversity order, meaning the number of effectively independent views of the chest. The theory predicts two regimes. The first is small-scale fading: multipath cancellation and micro-motion that changes from antenna to antenna across a co-located array. The combined signal-to-noise ratio of $M$ antennas under maximum ratio combining (weighting each antenna by its own channel before summing) has a coefficient of variation that shrinks roughly as $1/\sqrt{M_{\text{eff}}}$, where $M_{\text{eff}} \le M$ counts effectively uncorrelated antennas. This is the law-of-large-numbers argument behind hardening, described in What Will 5G Be?. The second is large-scale shadowing by the body itself, which is common to all elements of one array. We therefore model the undecidable fraction as

$$U(M) \approx U_{\text{shadow}} + \big(U(1) - U_{\text{shadow}}\big)\, \rho(M_{\text{eff}}),$$

where $U(M)$ is the fraction of windows rejected with $M$ antennas, $U_{\text{shadow}}$ is the fraction lost to body blockage that one array cannot see around, and $\rho$ is a decreasing function that the experiments will measure. The claim is falsifiable in three ways: $U(M)$ does not fall with $M$, it falls without plateauing, or distributing the same antennas across several viewpoints fails to lower the plateau.

The second half of the claim is statistical. Losing coverage biases estimates of AF burden, the fraction of time spent in AF, whenever losses are not random. If AF episodes coincide with restlessness or position changes, the rejected windows are enriched for AF. The naive burden computed from observed windows then underestimates the true burden. The error can be written as

$$\hat{B}_{\text{naive}} - B = \frac{B\,\pi_A}{B\,\pi_A + (1-B)\,\pi_N} - B,$$

where $B$ is true burden and $\pi_A$, $\pi_N$ are the probabilities that a window is observable during AF and during normal rhythm. More antennas help twice: they shrink the missing fraction, and they push $\pi_A$ and $\pi_N$ toward each other.

Relation to Prior Work

The closest work consists of three contactless cardiac systems, and all of them treat coverage as a nuisance. Atrial fibrillation detection via contactless radio monitoring and knowledge transfer trains a radar "student" network under an ECG "teacher" and detects AF from a 60 GHz radar. It requires stationary, isolated subjects and does not study array size. We reuse its detector design as the fixed classifier and vary only the sensing geometry. Monitoring long-term cardiac activity with contactless radio frequency signals (RF-HRV) extracts beat timing from high-order heartbeat harmonics without training. It gets beat timing right but reaches only 47.6% on complex arrhythmias, because timing alone cannot separate some regular abnormal rhythms from normal ones. That result motivates us to keep waveform shape and not just timing, and we include RF-HRV as a training-free baseline. MultiVital uses a small multi-antenna radar to separate five chest points. Its authors report that limited aperture blurs points at similar angles. That finding motivates larger apertures, but MultiVital measures signal fidelity, not coverage or diagnosis.

The large-array literature supplies the method. The analyses of maximum ratio combining and zero-forcing in What Role Can NOMA Play in Massive MIMO? and A Multi-Service Oriented Multiple-Access Scheme justify simple linear combining once arrays are large. The covariance eigen-decomposition of Cheng et al., which counts active users from dominant eigenvalues, gives us a cheap test for how many reflecting bodies are present in a shared-room condition. Those papers assume transmitters that send known pilot signals over synthetic channels. We apply the scaling argument to passive human reflectors whose "channel" is posture and breathing.

On the clinical side, sinus-rhythm AF prediction from ECG (Neal Yuan et al., Han et al.) shows that some underreporting can be sidestepped without catching an episode at all. We treat that as complementary and out of scope. Our target is capturing and quantifying episodes, which burden-guided decisions still require.

Research Program

We would build a bedside radar with a large virtual array, on the order of 100–200 virtual channels formed by several cascaded transmit and receive chips. Alongside it we would deploy three smaller units at different viewpoints (bedside, foot, ceiling) whose combined antenna count matches the large array. Every recording keeps raw per-channel data, so each array size and geometry can be simulated offline by selecting subsets of channels from the same night. This removes patient-to-patient and night-to-night variation from the scaling curves.

Participants would be patients with known paroxysmal AF, for example inpatients on continuous telemetry or patients shortly after catheter ablation, when episodes recur. All would wear a continuous multi-lead Holter ECG as ground truth, and a wrist optical pulse sensor as a wearable reference. A secondary condition adds a second sleeper to the room. Here we test whether covariance eigenvalue counting and zero-forcing separation keep per-person coverage from collapsing.

The analysis has three parts. First, for each subset size and geometry we measure the per-window signal-quality distribution and fit $U(M)$ and $U_{\text{shadow}}$. Second, we measure whether missingness depends on rhythm by comparing observability in ECG-confirmed AF windows with observability in normal-rhythm windows, estimating $\pi_A$ and $\pi_N$. Third, we fit a burden estimator that corrects for informative missingness. It reweights observed windows by their estimated observability given covariates the radar itself provides, such as body position and gross motion. We check this estimator against Holter burden.

Baselines

The do-nothing reference is a single antenna channel from the same recordings, passed through the same trained detector. This is what a minimal ambient sensor would see, and it fixes $U(1)$ together with the naive burden error.

The strongest existing method is the cross-modal teacher-student detector of Yuan et al.. That paper reports AF sensitivity 0.844 and specificity 0.995 on the RECG outpatient radar-ECG dataset of 6,258 visitors, plus test-retest reliability of 0.96 with 42.5 minutes of data. Those numbers come from seated or sleeping single subjects with good signal. The paper reports no coverage or undecidable fraction; no reported result was found on Lacuna for undecidable-window rates of contactless radar. We would reimplement the detector, train it once on pooled channels, and run the identical weights and rejection threshold on every array arm. Any difference then comes from sensing, not tuning. Beating it means the large array finds more of the ECG-confirmed episodes without lowering specificity below the 0.995 band. We would also show that combining channels, not the classifier, is responsible for the gain.

The training-free RF-HRV pipeline ( Monitoring long-term cardiac activity with contactless radio frequency signals) is run on the best single channel and on the combined array. It tests whether combining also rescues a timing-only method. Its reported 47.6% complex-arrhythmia accuracy is the reference point. The wrist sensor is scored with the 30-second windowing and quality rejection of Jeanningros et al., whose 32.4% undecidable rate is the published reference for wearable coverage.

Two ablations isolate the key ingredient. Best-antenna selection compares picking the single strongest channel per window with combining all channels at the same $M$. This separates true hardening from simply having more channels to choose among. The equal-antenna distributed arm tests whether gains come from diversity of viewpoint rather than aperture.

Why not just…?

Why not just record more nights with one cheap radar? If missed windows were random, longer recording would recover burden at low cost, and this is a real threat. It fails exactly when missingness is informative. Extra nights of a biased sampler converge to the biased value, $\hat{B}_{\text{naive}}$ above, not to $B$. Our estimates of $\pi_A$ and $\pi_N$ settle the question directly. If they are equal within confidence bounds, we will say that longer single-radar recording suffices for burden, and that arrays matter only for capturing short episodes.

Why not just several small radars around the room instead of one big array? This may well beat a co-located array. Body shadowing is common to one array, and viewpoint diversity is the only thing that removes $U_{\text{shadow}}$. That is why the equal-antenna distributed arm exists. If it dominates at every $M$, the practical conclusion becomes "distribute, don't enlarge." The scaling law still stands, with $M_{\text{eff}}$ governed by viewpoints rather than elements.

Why not just build a better classifier on a single radar? A stronger model raises accuracy on windows it keeps but cannot score windows where the heartbeat is buried. The fixed-detector design makes this explicit. We also report the single-channel detector's accuracy restricted to windows the large array deems decidable but the single channel rejects. If a better model recovers those windows from one channel, the hardware argument weakens accordingly.

Why not just use a smartwatch? Wearables lose windows too (32.4% in Jeanningros et al.). Timing-only analysis fails on flutter and frequent extra beats (65.4% and 65.9%). They also depend on nightly adherence, which Yuan et al. identify as a cause of missed episodes during sleep. The wrist arm quantifies this against the same Holter truth.

Decisive Checks

The project succeeds if three things hold on held-out patients. The undecidable fraction must fall with antenna count along a curve with a measurable plateau for the co-located array, and the plateau must drop in the distributed arm. The fraction of Holter-confirmed AF episodes of at least 30 seconds that are both observed and detected must rise with $M$ at fixed specificity. The missingness-corrected burden estimate must track Holter burden with smaller error than the naive estimate, particularly at small $M$.

The project fails informatively under either of two outcomes. If combining gives no advantage over best-antenna selection, hardening is not the mechanism. If rhythm-dependent missingness is absent, the statistical half reduces to a sample-size problem. Either outcome would tell designers of ambient monitors whether to spend on more antennas, more viewpoints, or simply more nights.