A blood analyser fires hundreds of thousands of cells through a laser and reports twelve numbers. A wearable does the same thing to a human being: it measures the body thousands of times a minute and reports a step count. This page tests what is still recoverable from the part that gets thrown away — on real patient data, held out, with the failures included.
The discarding is not an oversight — it is the design. A pedometer band-pass filters the accelerometer to the human walking cadence (~0.5–3 Hz) and counts peaks; actigraphy counts rectify and integrate a similar band (ActiGraph's classic filter passband is ≈0.25–2.5 Hz). Everything outside that band is, by definition, "not walking", so it is deleted. Everything inside an epoch collapses to one number.
Parkinsonian freezing of gait is trembling-in-place at 3–8 Hz. It falls outside the band and inside the epoch. It is deleted twice over.
Power spectrum of one ankle sensor over two four-second windows from a single recording. The device keeps the shaded cadence band and deletes the rest.
Three arms over identical 4-second windows. The only difference is how much of the signal survived the device. Every number below is measured on a patient the model never saw (leave-one-subject-out), with an identical standardised logistic regression on each arm.
The eight patients who had freezing episodes. Each line runs from what the device's own output achieves to what the raw stream achieves on that same patient.
The same experiment across four epoch lengths and two definitions of the summary arm — the faithful one, and a deliberately generous one that hands the summary waveform statistics no shipping device exposes. Raw wins in all eight. The margin is largest where temporal precision matters most, and narrows at coarse epochs, where the question being asked is easier but episodes can no longer be localised.
A model that reads a scan is graded on whether it survives the same patient being imaged on a different machine. That is one of the seven dimensions of MedEval-1, an independent medical-AI measurement standard: cross-site robustness. It is a formula, not a sentiment — train on one acquisition setting, test on the others, and score the fraction of chance-corrected skill that survives.
A wearable has the same problem in a different costume. The sensor sits somewhere else on the body; the cheap device samples more slowly. So we asked the sensor version of the question with the identical formula and the identical result format — and the discarded 3–8 Hz band is the only thing the model is allowed to see.
The number the standard tells you not to read alone. Judged on retention only, the sample-rate matrix scores higher than the placement matrix. It does not deserve to. The 8 Hz row keeps everything it has because it has almost nothing to lose — which is exactly the failure the standard's own §2.2.1 declares inadmissible, and exactly why it requires the absolute retained skill to be published beside every retention. Here that column reads 0.000: a model trained at 64 Hz and deployed on an 8 Hz device is at chance, with inverted ranking (AUC 0.31, κ −0.01). It does not degrade. It fails confidently, and silently.
The same method was run on a second modality — the wrist optical (PPG) sensor in PPG-DaLiA, 15 subjects, ~2.5 h each of free-living activity — asking whether the raw optical waveform can recover respiration rate, a core vital sign that no watch reports. Respiration is genuinely present in the raw waveform by three known mechanisms, and it is destroyed when the device collapses the waveform to a heart rate.
It did not work. The summary arm — given the ECG-derived ground-truth heart rate, a perfect measurement no shipping wearable achieves — predicted respiration rate better than the raw waveform did. Adding the raw waveform on top of the summary made it worse, not better.
What this says. Heart rate and respiration are both driven by metabolic demand, so a good heart rate is already a strong proxy — and extracting respiratory modulation from a wrist sensor under everyday motion is hard, which the literature has long held. The discarded 99% is not free signal everywhere. It pays where the deleted band carries a specific pathological signature the summary was designed to filter out — which is exactly the Parkinson's case above, and exactly the case that matters clinically.
What would falsify this. The Parkinson's result rests on 8 patients with freezing episodes in a lab protocol designed to provoke them. It has not been shown on free-living data, on a wrist-worn consumer device rather than research-grade ankle/thigh/trunk sensors, or against a clinical endpoint. Those are the next three experiments, not claims made here.