REMSleep

REM sleep detection for the Pebble 2 and Pebble Time 2, layered on top of Pebble Health.

Pebble Health auto-tracks sleep but has no REM concept and hides awake times. This app reads last night's Light/Deep sessions and per-minute heart-rate and movement data from Pebble Health, then carves out a REM stage and an awake timeline — viewable directly on the watch.

Platform

• Pebble 2 / Pebble 2 HR (diorite), with HR monitor
• Pebble Time 2 (emery), color, with HR monitor
• C SDK, health capability enabled

The hypnogram uses color constants; on the color Emery display they show as intended, and on the black-and-white Diorite display they render in B&W.

Building

Built on CloudPebble. To build with the local SDK instead:

pebble build
pebble install --emulator diorite   # or emery, or --phone <ip>

How it works

• health_service_activities_iterate() collects HealthActivitySleep (Light) and HealthActivityRestfulSleep (Deep) sessions over the last 24h.
• A per-minute array defaults to Awake; Light sessions mark Light, Deep overrides; gaps are Awake.
• health_service_get_minute_history() supplies per-minute HR + movement (vmc).
• REM is carved only from Light minutes using a Walch-feature-inspired weighted score (normalized HR, HR variability proxy, HR delta), gated by a REM latency window, a minimum episode length, and short-gap bridging.

Invariant

App Deep == Pebble Deep. App Light + REM == Pebble Light. App Awake == gaps.

Tuning

REM_SCORE_THRESHOLD (top of src/main.c) is the main dial — range 80–130, default 100.

Limits

• Pebble minute data lags ~15 min.
• HealthService API is read-only; results cannot be written back to Pebble Health.
• True HRV (RR intervals) and respiration are not exposed by the Pebble API, so minute-averaged HR is the ceiling. This is a heuristic, not the trained Walch model.

Reference

Walch et al. 2019, "Sleep stage prediction with raw acceleration and PPG heart rate" (Sleep, zsz180).

License

MIT — see LICENSE.