← Tepna preprints

Absent effort is not absent motion: chest-accelerometer respiratory effort fails to type apneas during CPAP therapy, and a mechanism for why

Michal Planicka  ·  corresponding author — Tepna Project

MotionDex (inertial motion) · Integrator (fusion) nodes, Tepna physiological-signal suite

Draft v1 · July 2026 · Data class: real-data, n-of-1, negative result · Analysis tool: resp-acc-analysis.html · 100% local

Abstract

Rationale. Distinguishing central from obstructive apnea rests on respiratory effort: absent in central events, preserved in obstructive ones. Wearable systems increasingly substitute a chest accelerometer for the polysomnographic effort belt, and the substitution is usually justified by asserting that effort amplitude collapses during a central event. Using a single effort channel, Nassi et al. recovered 84% of central apneas but only 51% of obstructive ones, and named the reason: "When using a single effort signal, thoracoabdominal asynchrony is undetectable" [8]. That asymmetry — good at detecting absence, poor at confirming presence — is the premise we test. Methods. 22 nights of Polar H10 chest accelerometry during CPAP therapy, time-aligned to the device's own AASM-typed scored events (392 apneas: 370 central, 31 obstructive with usable context). Respiratory-band effort envelope amplitude during each event was compared to three baselines (near, far, and night-level) under two windows. Results. Effort during central apneas sits at 0.99× the night-level baseline — not absent, but normal; obstructive events sit at 1.72×. The best discriminating configuration reaches AUC 0.691 (p = 0.0002), statistically robust but far below clinical utility. Framed as the literature frames it, only 16.5% of central apneas show amplitude below half baseline, where a single RIP belt achieves 84% recall [8]. An absolute amplitude threshold — the mechanism most wearable implementations use — would therefore mark effort "present" during the great majority of central events. A methodological finding of independent interest: including each event's termination in the analysis window destroys the effect entirely (AUC 0.573, p = 0.089), because the terminating arousal injects motion exactly where effort should appear absent. Interpretation. We tested the obvious explanation — that positive airway pressure mechanically drives chest-wall motion in the absence of respiratory effort — against the concurrent mask-pressure channel, and it fails: effort during central apnea is negatively associated with pressure (ρ = −0.174, p = 0.0008, n = 367), the opposite of the prediction. That narrows the explanation space to the transducer difference (a belt measures circumference, an accelerometer measures tilt) and to label quality (device-scored, not PSG). Every prior chest-accelerometer apnea study we identified recorded untreated diagnostic nights, so this failure mode remains untested elsewhere.

Keywords: central sleep apnea · respiratory effort · accelerometry · CPAP · negative result · treatment-emergent central apnea · AASM

1. Introduction

The AASM defines apnea by a ≥90% drop in airflow from baseline, and types it by effort: obstructive if effort persists, central if it does not. Effort is conventionally measured with thoracic and abdominal RIP belts. A chest accelerometer is an attractive substitute — it is already worn, it is cheap, and its respiratory-band amplitude is a plausible effort surrogate [1,6].

The evidence for the substitution is more asymmetric than it is usually presented. Nassi et al., using a single effort belt over 9,656 clinical and 8,455 SHHS recordings, classified 84% of central apneas correctly but only 51% of obstructive ones, and attributed the gap to the unobservability of thoracoabdominal paradox from one sensor [8]. Chang et al., using two accelerometers plus oximetry and ECG, reached 83.3% four-class event accuracy and stated explicitly that amplitude ratio is the central cue while paradox is the obstructive one [10]. Systems with a chest accelerometer plus a second modality report usable central detection — WatchPAT at sensitivity 67% / specificity 100% [11]; a chest-plus-abdomen accelerometric patch system at central-apnea-index r = 0.82 [12]. Notably, the most recent chest-accelerometer AHI system states in its own limitations that it "is not capable of distinguishing obstructive from central apnea events" [2].

What is missing from all of it: none of these recordings were made during positive-airway-pressure therapy.

2. Methods

22 nights of Polar H10 chest accelerometry, aligned to the concurrent ResMed recording by the cross-correlation and drift procedure described in the companion methods paper [9]. Events and their AASM types were read from the device's EDF+ annotation file. Only apneas of ≥8 s with ≥90 s of preceding context were retained: 370 central and 31 obstructive.

Three effort statistics were derived from the band-passed (0.13–0.50 Hz) tri-axial signal: the per-channel maximum, the median, and the projection onto the night's dominant principal component, each as a 1 s RMS envelope at 5 Hz with channels scaled to unit whole-night standard deviation. Each event's effort was expressed as a ratio to three baselines — a near window (−35 to −5 s), a far window (−90 to −60 s), and the night-level median — under two analysis windows: the full scored event, and its first 70% (excluding the termination). Discrimination was quantified by AUC with a Mann–Whitney test.

Why relative, not absolute. The AASM defines apnea and hypopnea as fractional drops from the patient's own recent baseline. Tilt-derived amplitude is additionally posture- and coupling-dependent, so an absolute milli-g threshold conflates physiology with sensor geometry. Both considerations require a relative measure; we report absolute-threshold behaviour only to characterise what a fixed-gate implementation would do.

3. Results

Effort stat | window | baselinemedian centralmedian obstructiveAUCp
max | first 70% | night0.991.720.6910.0002
PCA | first 70% | near1.313.550.6810.0004
median | first 70% | night0.972.050.6620.0013
max | full event | near1.361.730.5730.089

3.1 The effect is real but small. The best configuration reaches AUC 0.691 (p = 0.0002, n = 370 vs 31). Every "first 70%" variant outperforms its full-event counterpart across all three effort statistics and all three baselines, so the window effect is systematic rather than incidental.

3.2 The termination destroys the signal. Including the event's terminating arousal collapses discrimination to AUC 0.573 (p = 0.089). This is worth reporting on its own account: an analysis that windows on the scored event as annotated — the obvious implementation — measures the arousal, not the apnea.

3.3 Effort is not absent during central events. Central apneas sit at 0.99× the night baseline. Against the thresholds an absolute gate would use:

Thresholdcentral belowobstructive belowsingle RIP belt [8]
ratio < 0.34.6%3.2%
ratio < 0.516.5%6.5%84% central recall
ratio < 0.732.4%19.4%

A fixed-amplitude effort-present test would therefore mark effort present during 83.5–95.4% of central apneas, and type them obstructive.

4. Discussion — a hypothesis, not a conclusion

The discrepancy with Nassi et al. is large: 16.5% versus 84% recall of the effort-absent condition. Three explanations are available and are not distinguished by this data.

(a) Positive airway pressure drives chest motion without effort — TESTED, AND NOT SUPPORTED. Every study cited above recorded untreated diagnostic nights. Under CPAP the airway is pressurised and the machine continues to cycle, so chest-wall displacement might persist through a central event in which no respiratory effort occurs — which would produce exactly the observed 0.99× ratio. We tested this against the concurrently recorded MaskPress.2s channel on 398 events across 22 nights. The hypothesis fails, and the effect runs the other way. Effort during central apnea is negatively associated with concurrent pressure (Spearman ρ = −0.174, p = 0.0008, n = 367; Pearson r = −0.123, p = 0.018); splitting at the median pressure, low-pressure events carry a median ratio of 1.12 against 0.83 at high pressure. If pressure were mechanically driving the residual motion, the association would be positive.

How much this settles. Less than it appears. Pressure varied only over the therapeutic range actually delivered (4.14–11.08 cmH₂O, IQR 5.65–7.79), so this is a within-therapy dose–response test, not a therapy-versus-no-therapy comparison. A null — or even a reversed association — inside that range does not exclude the possibility that being on therapy at all raises the motion floor relative to an untreated night. The asymmetry matters: a positive result would have been strong evidence for the hypothesis, whereas this negative is only moderate evidence against it. Settling it properly requires an untreated night, which an all-CPAP corpus cannot supply.

(b) Transducer difference. A RIP belt measures circumference; an accelerometer measures tilt. These are not the same quantity and need not fail together.

(c) Label quality. The event types are the CPAP manufacturer's algorithm, which infers central events from forced-oscillation airway patency, not expert AASM scoring against esophageal pressure. A recent expert review notes that accelerometric chest-movement correlation with esophageal pressure "remains insufficiently characterized" [13].

5. ⚠ Status and limitations

This paper is PARKED and should not be cited as establishing its title claim. Four issues must be resolved first: (i) the CPAP-pressure hypothesis (§4a) has now been tested and fails, so the convenient narrowing to "under therapy" is not available — but the within-therapy design means the mechanism behind the 0.99× ratio remains genuinely unexplained, and publishing an unexplained negative is weaker than publishing an explained one; (ii) obstructive n = 31, so the AUC confidence interval is wide and the obstructive median is unstable; (iii) the labels are device-scored, not PSG; (iv) an independent adversarial review of the supporting literature found a prior report whose direction may run opposite to the mechanism assumed here, and that contradiction is unresolved.

Beyond those: one subject; one posture (gravity-roll IQR 13.1°–17.9°, so postural modulation of effort coupling is unexplored); and the effort envelope, though band-limited to 0.13–0.50 Hz with a zero-phase 4th-order Butterworth measured at −36.6 dB at 0.8 Hz, cannot be guaranteed free of all arousal-related motion even in the first 70% of an event.

What the result does support, and we believe robustly, is the narrow engineering claim: on this corpus, an absolute-amplitude effort-present test does not type apneas, and any implementation relying on one should abstain rather than emit a type.

6. Reproducibility

⛔ Status. resp-acc-analysis.html now exists and reproduces the original harness on spot-checked nights; the full-corpus regeneration through it is still outstanding, so this paper remains a DRAFT. See briefs/MOTIONDEX-RESPIRATORY-RATE-2026-07-21-BRIEF.md §4.

7. References

Citation provenance. verified directly by the author against the publisher record, including quoted statistics. retrieved by automated literature agents with full text obtained, not independently re-verified. full text never obtained.
  1. Schipper F, van Sloun RJG, Grassi A, Derkx R, Overeem S, Fonseca P. Estimation of respiratory rate and effort from a chest-worn accelerometer using constrained and recursive principal component analysis. Physiol Meas 42(4):045004, 2021. doi:10.1088/1361-6579/abf01f
  2. Schipper F, Fonseca P, Grassi A, Ross M, et al. Apnea-hypopnea index estimation using overnight chest-wall accelerometry. Front Sleep, 2026. doi:10.3389/frsle.2026.1858267 — source of the verbatim limitation quoted in §1.
  3. Charlton PH, Bonnici T, Tarassenko L, Clifton DA, Beale R, Watkinson PJ. An assessment of algorithms to estimate respiratory rate from the electrocardiogram and photoplethysmogram. Physiol Meas 37(4):610–626, 2016. doi:10.1088/0967-3334/37/4/610
  4. Ryser F, Hanassab S, Lambercy O, Werth E, Gassert R. Respiratory analysis during sleep using a chest-worn accelerometer: a machine learning approach. Biomed Signal Process Control 78:104014, 2022. doi:10.1016/j.bspc.2022.104014
  5. Doheny EP, Lowery MM, Russell A, Ryan S. Estimation of respiration rate and sleeping position using a wearable accelerometer. Proc IEEE EMBC, 2020:4529–4532. doi:10.1109/EMBC44109.2020.9176573
  6. Schipper F, van Sloun RJG, Grassi A, Overeem S, Fonseca P. A deep-learning approach to assess respiratory effort with a chest-worn accelerometer during sleep. Biomed Signal Process Control 83:104726, 2023. doi:10.1016/j.bspc.2023.104726
  7. Akbarian S, Montazeri Ghahjaverestan N, Yadollahi A, Taati B. Distinguishing obstructive versus central apneas in infrared video of sleep using deep learning. J Med Internet Res, 2020. doi:10.2196/17252
  8. Nassi TE, Ganglberger W, Sun H, et al. Automated scoring of respiratory events in sleep with a single effort belt and deep neural networks. IEEE Trans Biomed Eng, 2022. doi:10.1109/TBME.2021.3136753 — source of the 84%/51% asymmetry and the quoted sentence.
  9. Planicka M. The machine in the bedroom: CPAP flow as a home reference standard. Tepna working preprint, 2026. cpap-flow-reference.html
  10. Chang HC, Wu HT, Huang PC, Ma HP, Lo YL, Huang YH. Portable sleep apnea syndrome screening and event detection using LSTM recurrent neural network. Sensors 20(21):6067, 2020. doi:10.3390/s20216067
  11. Pillar G, Berall M, Berry R, et al. Detecting central sleep apnea in adult patients using WatchPAT — a multicenter validation study. Sleep Breath, 2020. doi:10.1007/s11325-019-01904-5
  12. Rohrscheib C, Moura AA, Raphelson J, Orr JE, Patel RP, Malhotra A. Evaluation of an automated sleep apnea scoring system using body-worn effort patches. Sleep Med, 2026.
  13. Chiang AA, et al. From classic to cutting-edge: technological approaches to respiratory physiological signals in sleep medicine. J Clin Sleep Med, 2026.

Tepna is not a medical device and is not intended to diagnose, treat, cure or prevent any disease. Research and personal-analysis use only.

v2.8.0