PillowGraph · drawing set · v1.0

Part I · Cover Docket · Abstract · v1.0 skeleton

U.S. Provisional Patent Application · Drawing set

Appl. No. —

Sheets 5

Rev. B · v1.0

Title · Method for sleep-stage inference from a hand-held computing device positioned within a pillow envelope, using its inertial-measurement-unit and microphone with on-device event-token reduction

Inventor · J. H. Kim

Internal · PillowGraph

Filing target · pending overnight power-budget gate

Classification (proposed) · A61B 5/00 · A61B 5/4806

About this Personal project · v1.0 skeleton Gate-pending overnight test

Draft cover sheet and drawings for PillowGraph. The W1 gate for this card is a thermal / battery test of an iPhone wedged in a pillowcase over a single eight-hour night. If the phone survives the night without exceeding 40 °C surface temperature or 40 % battery drain, the gate clears and the card upgrades from v1.0 skeleton to a full provisional draft. v1.0 honest acknowledgment (2026-05-18): the headline mechanic — phone IMU + microphone → on-device sleep-stage classification — is squarely anticipated by SleepCycle US 8,493,220 B2 (Smart Valley Software / Northcube; phone microphone → sleep stages on-device, with light-sleep alarm) and by the Pillow app (iPhone IMU + microphone → REM / Light / Deep stages, commercial since 2014). Earlier revisions treated this as the contribution; v1.0 demotes it to acknowledged prior art and narrows the claim to the genuinely-surviving combination: phone inside the pillow envelope (mechanical coupling, not nightstand-acoustic), event-token reduction before any audio is persisted to non-volatile storage, and the specific snore / stir / wake / baseline discrete vocabulary.

Abstract of the disclosure Cover-sheet boilerplate v1.0

A system and method infer sleep stages from a commodity hand-held computing device wedged within a pillow envelope while a user sleeps thereon. The device's built-in inertial-measurement unit (IMU) captures sub-millimetre head and respiratory motion through the pillow material; its built-in microphone captures ambient acoustic events. Acoustic samples are classified locally into a small token vocabulary — snore, stir, wake, baseline — before persistence, such that no raw audio is retained on-device beyond the classification window. The IMU stream and event-token stream jointly drive a four-stage sleep-stage classifier (Wake / REM / Light / Deep) that produces a single morning summary visualisation. No raw sensor data or audio leaves the device.

Field

A61B 5/4806 · sleep-stage detection

Acknowledged prior art (headline mechanic)

SleepCycle US 8,493,220 B2 — phone mic → sleep stages on-device
Pillow app (iOS, 2014–) — iPhone IMU + mic → REM/Light/Deep
Apple Beddit (2017–) — under-mattress sensor strip

Distinguished from (surviving novelty)

SleepCycle places phone next to bed · this discloses phone IN pillow envelope
Pillow app retains audio · this performs event-token reduction before persistence
My Pillow Knows My Sleep — computational fabric, not commodity phone
Apple Watch / Oura / Withings — different sensor surface (wrist / finger / mat)

Index of sheets Tap a row to jump 5 sheets

Sheet 1 FIG. 1 Pillow cross-section · phone + acoustic cone 100
Sheet 2 FIG. 2 8-hour hypnogram + event tokens 200
Sheet 3 FIG. 3 Pipeline · IMU + mic → tokens → stage 300
Sheet 4 FIG. 4 Event-token privacy boundary 400
Sheet 5 FIG. 5 Overnight power + thermal budget · W1 gate 500

Part II · Drawings FIG. 1 – 5 · Sheets 1 – 5

Sheet 1 / 5 Representative FIG. 1 · Pillow cross-section · phone + acoustic cone 100

FIG. 1

Sheet 2 / 5 FIG. 2 · 8-hour hypnogram + event tokens 200

FIG. 2

Sheet 3 / 5 FIG. 3 · Pipeline · IMU + mic → tokens → stage 300

FIG. 3

Sheet 4 / 5 FIG. 4 · Event-token privacy boundary 400

FIG. 4

Sheet 5 / 5 FIG. 5 · Overnight power + thermal budget · W1 gate 500

FIG. 5

Part III · Specification Background · Summary · Brief description

Background of invention Prior-art context

Existing consumer sleep-tracking products fall into three categories: (i) wrist-or-finger wearables (Apple Watch, Oura, Fitbit), (ii) under-mattress or under-pillow dedicated sensors (Withings Sleep, Beautyrest, Apple Beddit; computational-fabric pillowcases per UbiComp '25's My Pillow Knows My Sleep), and (iii) phone-based microphone-plus-IMU apps using the device's own sensors (SleepCycle, Pillow). The wearable category requires a worn device; the dedicated-sensor category requires specialised hardware purchase; the phone-app category is the closest art to this disclosure and is treated explicitly below.

The closest patent reference is SleepCycle's US 8,493,220 B2 (Virtanen, Salmi, Salmi; assigned to Smart Valley Software, later Northcube AB), which discloses a mobile-device microphone capturing sound signals from a sleeping subject's movements, classifying sleep stages on-device from those signals, and detecting arousal/awakening for light-sleep alarming. The closest commercial reference is the Pillow app (iOS, 2014 –), which fuses iPhone IMU (accelerometer + gyroscope) and microphone on-device to produce REM / Light / Deep sleep-stage estimates. Earlier revisions of this page treated the phone-mic + phone-IMU → on-device sleep-stage pipeline as the contribution; v1.0 demotes that to acknowledged prior art. Apple's Beddit family (acquired 2017; extended in patents to smart-mattress and slim-bed-sensor embodiments through 2024) covers under-mattress sensor strips and is distinct in deployment geometry from the phone-in-pillow form factor disclosed here.

The disclosed system survives the foregoing as a combination: (i) phone IS positioned inside the pillow envelope rather than on a nightstand — SleepCycle's reference embodiment places the phone "next to" the bed, with sensitivity to room reverberation that the in-pillow geometry materially avoids by direct mechanical coupling through the pillow material; (ii) audio is reduced to a finite event-token vocabulary of snore / stir / wake / baseline BEFORE any persistence — the Pillow app and SleepCycle retain audio (Pillow expressly offers "audio recording" as a user-toggleable feature) whereas this system commits to never writing raw audio to non-volatile storage; and (iii) the temporal classifier (FIG. 4) consumes the discrete token sequence rather than raw acoustic features, which makes the privacy guarantee structural rather than policy-based.

Summary of the invention per 37 CFR § 1.73

The disclosed system infers sleep stages from a commodity hand-held device positioned within the pillow envelope, with no auxiliary sensor (no wearable, no under-mattress sensor, no bedside microphone), and with the acoustic stream reduced to a discrete event-token vocabulary before any persistence — such that no raw audio sample of the night is ever written to non-volatile storage on the device.

A hand-held computing device 104 positioned within a pillow envelope 102 reads IMU stream 302 and microphone stream 304 across a sleep window. Motion-feature extractor 306 produces respiration and gross-motion features from the IMU; acoustic classifier 308 reduces the microphone stream to discrete event tokens (snore, stir, wake, baseline) on a windowed basis. Fusion stage 310 combines features and tokens into a four-stage hypnogram 312 via a hidden-Markov model or analogous temporal classifier. Raw audio samples are not retained beyond the per-window classification.

The output is a single morning visualisation (FIG. 2 style) intended for ambient consumption rather than dashboard interaction.

Brief description of drawings Sheets 1 – 4

FIG. 1Side cross-section of the bed scene: pillow envelope 102 resting on mattress 112 (hatched) under sheet 114; head 108 with respiration 116 exiting the airway; phone 104 (host computing device) embedded within the pillow, coupled to the head via mechanical-coupling region 118; IMU axes 106 (x · y · z); mic acoustic cone 110 aimed at the airway. Scale bar 120 indicates a ≈ 50 cm queen-pillow span.
FIG. 28-hour hypnogram with four stage lanes — Wake 202, REM 204, Light 206, Deep 208 — at 80 px/hour linear time scale (23:00 → 07:00). Hypnogram polyline 216 traces inferred stage; event tokens (snore 210, stir 212, wake 214) populate the upper margin with audio reduced to discrete events on-device. Night summary 218 reports total sleep time, efficiency, REM %, and deep %.
FIG. 3Two-lane on-device pipeline 300: IMU lane 322 (100 Hz, hatched-hw block 302 → motion features 306) and MIC lane 324 (8 kHz, hatched-hw block 304 → acoustic classifier 308). Both lanes converge at fusion + HMM 310; output hypnogram 312 routes to morning view 314. Bottom strip 326 lists clock domains and battery budget; legend 328 distinguishes hardware / software / user surface.
FIG. 4Privacy-boundary diagram: on-device region 402 holds raw IMU 406, raw audio 408, motion features 410, event tokens 412, and hypnogram 414 (≈ 460 MB / night retained locally); off-device region 404 (hatched) lists possible destinations 416 — cloud sync, family sharing, clinician export, research opt-in. Only hypnogram 414 may cross the boundary, user-gated (≈ 1 KB). Data-budget strip 418 reports the retained-to-egress ratio (≈ 460 000 : 1).
FIG. 5Overnight power + thermal budget visualising the W1 gate (Part VI condition): battery % 502 declining from ≈ 95 → 68 (drain 27 %) and surface temperature 504 rising from ≈ 26 → 36 °C over 23:00 → 07:00, both staying below the battery floor 506 (≤ 40 % drain) and the temperature ceiling 508 (≤ 40 °C). PASS zone 510 shaded; verdict strip 514 reports gate-open, eligible for v1.0 → provisional promotion.

Part IV · Claims 5 total · 1 indep · 3 dep · 1 apparatus

Claims 1 independent · 3 dependent · 1 apparatus Draft v1.0

What is claimed is:

1. A method for inferring sleep stages, the headline mechanic of on-device phone-microphone + phone-IMU sleep-stage classification being acknowledged in the prior art (SleepCycle US 8,493,220 B2; Pillow app, iOS 2014–), comprising:

(a)positioning a hand-held computing device (104) within a pillow envelope (102) upon which a user sleeps;
(b)capturing, at said hand-held computing device, an inertial-measurement-unit stream (302) and a microphone stream (304) across a sleep window;
(c)reducing said microphone stream, on-device, to a sequence of event tokens drawn from a finite vocabulary including at least snore, stir, wake, and baseline (308);
(d)discarding raw audio samples of said microphone stream beyond the per-window classification employed in step (c);
(e)computing a four-stage hypnogram (312) from said inertial-measurement-unit stream and said event-token sequence by a temporal classifier;
(f)presenting said hypnogram via an ambient visualisation at a user-configurable time of day; and
(g)wherein no sensor external to said hand-held computing device is recruited for said inferring of sleep stages, the audio reduction of (c) is performed before persistence such that no raw audio sample of said microphone stream is ever written to non-volatile storage on said device, and no raw audio, no raw inertial-measurement-unit sample, and no per-user identifier is transmitted off-host.

Survives over: SleepCycle US 8,493,220 B2 (phone next to bed, not in pillow envelope; audio retained on device) and Pillow app (audio expressly recorded as user feature). Surviving novelty carried by the in-pillow positioning in (a), the discrete-vocabulary event-token reduction in (c), and the structural no-raw-audio-to-storage guarantee in (d) + (g).

2. The method of claim 1, wherein the inertial-measurement-unit stream is sampled at a rate of 50 to 200 Hz and the microphone stream at 8 to 16 kHz.

3. The method of claim 1, wherein the temporal classifier comprises a hidden-Markov model trained on a representative population sleep dataset prior to deployment.

4. The method of claim 1, wherein presentation of the hypnogram is implemented as a single non-interactive ambient graphic, not as a metric dashboard.

5. A hand-held computing device comprising an IMU, a microphone, one or more processors, and non-transitory memory storing instructions which, when executed, cause the device to perform the method of claims 1 – 4.

Claims · 5 total · 1 independent · 3 dependent · 1 apparatus

Part V · Appendices Prior-art bibliography

Prior-art bibliography Selected; not exhaustive

Closest art (v1.0 audit add)
Virtanen, V., Salmi, A. and Salmi, S. (Smart Valley Software / Northcube AB). Method for determining stages of sleep using a personal mobile device. US 8,493,220 B2 (granted 2013). [SleepCycle's foundational patent · phone mic + on-device sleep-stage classification; reference embodiment places phone next to bed, distinct from in-pillow geometry]
Neybox Digital Ltd. Pillow — Sleep Tracker. iOS app, 2014–. [closest commercial reference · iPhone IMU + microphone → REM/Light/Deep; retains audio as a user-toggleable feature, distinguishable from this system's pre-persistence event-token reduction]
Apple Inc. / Beddit Oy. Sleep monitoring patent family extending the original Beddit sensor strip (Apple acquired Beddit May 2017; family updated through 2024 to smart-mattress and slim-bed-sensor embodiments). [different deployment geometry · under-mattress sensor not in-pillow phone]
Adjacent art
Shao, Q. et al. My Pillow Knows My Sleep: Sleep Monitoring with Computational Fabrics in the Pillowcase. UbiComp 2025.
Behar, J. et al. SleepAp: A Smartphone-based App for Sleep Apnea Screening. J. Med. Internet Res. 2013.
Withings SA. Sleep tracking mat product line. 2017–.
Rechtschaffen, A. and Kales, A. A manual of standardized terminology, techniques and scoring system for sleep stages. 1968 (R&K hypnogram standard).
Iber, C. et al. The AASM Manual for the Scoring of Sleep and Associated Events. 2007 (modernised scoring).

Part VI · Execution Version · v1.0 · Skeleton

Version history Draft · not filed

v1.02026-03-25 · Skeleton draft. Gate-pending the single-night thermal/battery test.

Promotion to a full provisional draft is conditioned on a single overnight experiment confirming acceptable phone thermal and battery behavior with the device wedged in a pillowcase for an eight-hour window.

/pillowgraph · v1.0 · drawing-stage

Index