MotionDex (inertial motion) · Integrator (fusion) nodes, Tepna physiological-signal suite
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
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.
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.
| Effort stat | window | baseline | median central | median obstructive | AUC | p |
|---|---|---|---|---|
| max | first 70% | night | 0.99 | 1.72 | 0.691 | 0.0002 |
| PCA | first 70% | near | 1.31 | 3.55 | 0.681 | 0.0004 |
| median | first 70% | night | 0.97 | 2.05 | 0.662 | 0.0013 |
| max | full event | near | 1.36 | 1.73 | 0.573 | 0.089 |
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:
| Threshold | central below | obstructive below | single RIP belt [8] |
|---|---|---|---|
| ratio < 0.3 | 4.6% | 3.2% | — |
| ratio < 0.5 | 16.5% | 6.5% | 84% central recall |
| ratio < 0.7 | 32.4% | 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.
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.
(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].
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.
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.Tepna is not a medical device and is not intended to diagnose, treat, cure or prevent any disease. Research and personal-analysis use only.