A battery-free radio sensor can watch someone breathe. When a person sits near a cheap energy harvester that is soaking up ambient Wi-Fi, their chest motion slightly changes how much radio energy reaches it, and the harvested DC voltage rises and falls with each breath. REHSense showed this works at about 12 mW, roughly 98.7% less power than conventional Wi-Fi channel-state sensing. Its respiration estimator, though, is a hand-built peak counter, and its reported 95.4% accuracy is for subjects lying down. The obstacle to doing better is labels. A learned model that copes with sitting, standing, fidgeting or a second person in the room needs a clean reference breathing signal for every training window, which means wearing a reference sensor or having an expert score the trace. Unlabeled harvester recordings, by contrast, cost almost nothing to collect.
We propose to train a deep breathing estimator on unlabeled harvester streams, with no reference labels at all. Supervision comes from a small committee of classical signal-processing detectors, such as peak counting, spectral peak picking and autocorrelation period estimation. Each detector is reliable in some conditions and wrong in others. Our bet is that a student network can learn, from its own internal representation, which detector to trust on which window. This is the hierarchical "mixture of weak teachers" recipe from Co-Supervised Learning, carried from image classification to physical-sensor time series. If it works, the student should beat every individual detector in the hard regimes where none of them is reliable, and it should do so without an expert-labeled training set.
Core Claim
The detectors make different, regime-dependent mistakes. Peak counting picks up spurious peaks when posture changes add slow drift. Spectral methods return harmonics when breathing is irregular. Autocorrelation breaks down at low signal-to-noise ratio, for example at the far edge of the roughly 1.75 m sensing range REHSense reports. Because the errors depend on the regime, choosing a teacher per window should beat averaging the teachers per window. The claim is falsifiable in one direct comparison. If a student trained on the per-window median of all detectors matches the student trained with learned teacher selection, then the mixture-of-experts structure contributes nothing, and the proposal reduces to ordinary pseudo-labeling.
Relation to Prior Work
The method comes from Co-Supervised Learning, abbreviated CSL. CSL treats a set of specialized weak teachers as a hierarchical mixture of experts. It alternates between using the student's representation to assign each sample to its most competent teacher and training on that teacher's labels. It also discards samples where the teacher's label and the student's prediction disagree. CSL reports more than 15% gain in performance gap recovery over a single teacher on the OpenAI weak-to-strong benchmark, and 14–17% on DomainNet. Its teachers, however, are trained neural classifiers on images with discrete labels, and their specializations are defined by data domain. Our teachers are training-free algorithms producing a continuous rate and a breath-phase signal, and their specializations are defined by physical conditions. We therefore need to adapt CSL's assignment and filtering steps to regression, and to test whether its premise still holds when the teachers are this narrow.
Two other papers motivate the setting. REHSense supplies the sensing setup and our weakest-but-honest baseline. It also documents the gap we target: its learned models depend on 15,000 manually labeled segments, and it leaves multi-person interference unresolved. The review of leadless pacemakers is a second example of regime-shifting backscatter channels: received power swings by 3.1 dB between systole and diastole. We treat it as evidence that this structure is general, not as a test bed, because we have no access to implant data.
UMQ's quality-aware mixture of experts sends inputs to different sub-networks according to an estimated signal quality. Removing that routing costs 2.5 Acc2 points on degraded CMU-MOSI. This is the closest design to routing by signal regime. The difference is that UMQ routes features inside a supervised model, while we route supervision when labels are absent. UMQ also reports that its gating is sensitive to hyperparameters, so we keep the student itself dense.
Two papers constrain the design. On label-scarce physiological time series, the Ward2ICU study of synthetic vital signs found that GAN-generated proxy data nearly matched real data on common outcomes but fell from 47.8% to 38.0% accuracy on the rare ICU-transfer class. That result argues against substituting synthetic data. The scribble-supervision work shows that sparse labels combined with uncertainty filtering come within 2.9–4.5 Dice points of full supervision in medical segmentation. It supports the filtering component, but it still requires a human for every scan, which we try to avoid.
Research Program
The first component is the sensing data. We will use the REHSense breathing dataset ($D_{RM}$) if its authors release it: 300 minutes, 10 subjects, three postures, with FLOWTIME headband reference. The file does not state whether it is public. Otherwise we reproduce the setup with the same Powercast P21XXCSR-EVB harvester and FLOWTIME headband. We then add two kinds of recording. The first is a large unlabeled corpus, tens of hours, in which the headband is worn only so that it can be held out for evaluation. The second is a set of systematically varied conditions: three postures, distances of 0.5, 1.0 and 1.5 m, deliberate small movements, and a second person moving nearby.
The second component is the teacher committee. It contains the REHSense variance-peak counter, a band-limited FFT peak picker, an autocorrelation period estimator, and a zero-crossing counter on the Savitzky-Golay-filtered signal. Each teacher outputs a breathing rate and a binary breath-phase sequence for each 30-second window.
The third component is the student: a 1D convolutional network over the filtered voltage window, with a rate head and a phase head. Training follows CSL. Teacher assignment is driven by the student's embeddings: we cluster the embeddings, and within each cluster we assign the teacher that agrees most often with the student's current predictions. Consistency filtering discards windows where the assigned teacher's rate differs from the student's prediction by more than a tolerance, or where the window-local and cluster-level teacher choices disagree.
We will also produce an analysis, not only a model. Using the held-out reference, we compute which teacher was actually best in each regime, and we check whether the student's learned assignment recovers that map. If the assignment is no better than chance against that map, any accuracy gain is coming from something other than teacher selection.
Baselines
The do-nothing reference is the best single classical detector applied directly, with no learning. In practice this is the REHSense variance-peak counter, which reports 95.4% respiration accuracy for lying-down subjects. The file gives no REHSense respiration numbers for sitting or standing; no reported result found on Lacuna for those conditions. We will therefore re-measure the peak counter ourselves in every regime, on the same windows used for every other method. It is a strong competitor because it needs no data and already works well at rest. Any learned method has to beat it where it matters, in the non-resting, far-distance and interference regimes.
The strongest learned competitor is a student with the same architecture trained on the per-window median of all four detectors, with the same agreement-based filtering. This is the obvious simpler aggregation, and it isolates the key ingredient of our proposal: choosing a teacher per regime rather than averaging teachers. A second ablation is CSL without consistency filtering. Further reference points are the single-teacher pseudo-label student, which follows the single-generalist setup CSL improves on, and an oracle-assignment student that sees the reference only to pick the best teacher per window. The oracle bounds what teacher selection can achieve.
The ceiling is a fully supervised student trained on headband labels. A few-label variant uses 10 reference-labeled minutes per subject. It tests whether a small amount of expert-style labeling simply beats the whole weak-teacher pipeline.
All learned methods share the same unlabeled corpus, architecture, optimizer and number of training steps, and the same tuning budget of 20 configurations. Tuning is done on a reference-labeled development split of two held-out subjects, which no method uses for training. Evaluation is leave-subject-out on the remaining subjects. Metrics are mean absolute rate error in breaths per minute, REHSense-style accuracy for comparability with the 95.4% figure, phase F1 against the headband, and a performance gap recovery score adapted from CSL. That score is the fraction of the gap between the best single teacher and the supervised ceiling that a method closes.
We count the proposal as succeeding if it beats the median-fusion student by a margin larger than the across-seed standard deviation, beats the best single detector in at least the sitting, standing and 1.5 m regimes, and reaches at least half of the gap recovered by the oracle-assignment student.
Why not just…?
Why not just use the peak counter, which already gets 95.4%? That number is for subjects lying still. REHSense needed a learned classifier for anything dynamic, and it lists multi-person interference as unsolved. The regime-by-regime comparison will show directly whether the peak counter degrades off-rest. If it does not, there is nothing for a student to learn, and we will report that as the result.
Why not just fuse the detectors by median or vote and train on that? This is the real threat. If the detectors' errors are independent and mostly small, the median cancels them, and teacher selection adds nothing. Selection only helps when, in some regime, most teachers fail in the same direction and one teacher stays correct. Harmonic doubling, which affects both spectral and autocorrelation methods together, is a plausible case of this. The median-fusion baseline is designed to settle the question. Our per-regime breakdown will also show where, if anywhere, the two diverge.
Why not just label ten minutes per subject and fine-tune? Wearing a reference sensor briefly is cheap in a lab, so this alternative is fair. It does not scale to deployment, where each new home, distance and second occupant shifts the signal and would require relabeling. The few-label baseline measures this trade-off directly. If it dominates the weak-teacher student, the honest conclusion is that a little labeling wins at this scale.
Why not just synthesize labeled traces? The Ward2ICU results show that synthetic physiological series lose the rare, irregular cases: ICU-transfer accuracy fell from 47.8% to 38.0%. Irregular breathing and interference are exactly the cases we care about, so we do not pursue synthesis.
Decisive Checks
Three results decide the outcome. The first is the student-versus-median comparison in the hard regimes. The second is whether the student's learned teacher assignment agrees with the oracle per-regime best-teacher map. The third is whether the student exceeds the best single detector at 1.5 m and with a second person present. If all three hold, we have shown that classical detectors can replace expert labels for battery-free RF vital-sign sensing, and that the mixture structure is what makes it work. If only the third holds, then the gain comes from pseudo-label training in general, not from expert selection, and we will say so.
References
- REHSense: Towards Battery-Free Wireless Sensing via Radio Frequency Energy Harvesting
- Co-Supervised Learning: Improving Weak-to-Strong Generalization with Hierarchical Mixture of Experts
- Leadless pacemakers: A review of communication methods, energy management, and clinical applications
- Addressing Missing and Noisy Modalities in One Solution: Unified Modality-Quality Framework for Low-quality Multimodal Data
- Ward2ICU: A Vital Signs Dataset of Inpatients from the General Ward
- Learning to Segment Medical Images with Scribble-Supervision Alone