Skip to content

cva-cardiovascular-age.md

Latest commit

 

History

History
64 lines (51 loc) · 3.28 KB

cva-cardiovascular-age.md

File metadata and controls

64 lines (51 loc) · 3.28 KB
Raw

Cardiovascular age from raw PPG (cva_ppg → cva_raw_ppg_data → CVA model)

Models are not included in this repo. The decrypted cva_2_1_0.pt (and every other .pt) is Oura's proprietary IP; it is not committed (gitignored under notes/models/). Run this against your own locally-decrypted copy.

When the cva_ppg feature (CAP_CVA_PPG_SAMPLER, id 0x0d) is enabled, the ring records short raw-PPG bursts overnight and emits them as cva_raw_ppg_data events (BLE tag 0x81). This is the raw photoplethysmography waveform the cardiovascular-age (CVA) model needs — the one signal the PPG-gated models used to be blocked on. With it captured, we can run Oura's own CVA model offline.

Decoding tag 0x81

Each event body is signed int8 deltas; a cumulative sum reconstructs the PPG ADC samples. The Rust decoder (oura-protocol, decode_cva_raw_ppg) emits {"ppg_samples": [...], "n": N} per event (per-event cumsum). A full measurement spans a burst of consecutive events <2 s apart (~1503 samples ≈ 10 s at the ring-reported rate, ~140 Hz on Ring 5); concatenate the burst's deltas, then cumsum.

Validation that the decode is correct: the reconstructed waveform is pulsatile and its pulse rate matches the independent IBI-derived heart rate (e.g. 38 bpm from PPG peaks vs 41 bpm from IBI in the same window).

The model (cva_2_1_0, CardiovascularAgeV2Model)

forward(ppg_segments [n, 1500] f32, demographics [1, 5] f32)
  -> (daily_cva, quality, raw_quality, daily_pwv, ppg_segment_metrics[n, 11])
  • ppg_segments: groups of exactly 1500 raw samples (the app drops any group that isn't 1500). No filtering/normalization in the app — preprocessing is inside the .pt. Feed raw reconstructed samples.
  • demographics = [sex, height_m, age_years, ring_size, weight_kg], where sex = −1 female / +1 male / 0 other. These come from the user profile, not the ring (the ring's user_information event has no anthropometrics).

Outputs

  • daily_cva = vascular age (years). It is anchored to the input age and adjusted by PPG morphology (age 25 → 16.0, 40 → 24.6, 60 → 38.4 on our data — i.e. consistently ~10–15 y younger than chronological = healthy arteries). Pass real demographics for a meaningful absolute number.
  • daily_pwv = pulse-wave velocity (m/s), PPG-only (~4.5 m/s on our data; lower = less arterial stiffness).
  • quality / raw_quality = signal-quality scores; ppg_segment_metrics = 11 per-segment features.

Running it

python tools/run_cva_model.py [DB] --sex M --age 30 --height 1.78 --weight 75 --ring 10

On one night (10 valid 1500-sample segments from 18 measurements) with placeholder demographics: vascular age ≈ 18.8 y, PWV ≈ 4.5 m/s, quality 2 / raw 0.70.

Caveats

  • Absolute daily_cva is only meaningful with the user's real demographics.
  • We segment by a 2 s event-gap heuristic; the app groups by the protobuf measurementStartTimestamp (decoded by the native lib we don't run), so a borderline burst could be split differently. The model still produces a stable value.
  • The exact training sample rate isn't confirmed from Java; segments are fed as raw samples (the model bakes in its canonical rate). See tools/run_cva_model.py.