Skip to content

Use Cases & Value

compressionKIT helps wearable teams keep more waveform data without paying the full memory, radio, and cloud-ingestion cost of raw sampling. This page summarizes where compression usually creates value and gives order-of-magnitude sizing examples for product planning.

Who benefits

  • Consumer wearables

    Smartwatches, rings, patches, earbuds — any device with a PPG or ECG sensor and a tight battery / flash budget.

  • Remote patient monitoring

    Multi-day Holter patches, hospital-at-home kits, post-op monitors where storage and cellular cost dominate the BOM.

  • Sleep & wellness

    Overnight PPG for HR, HRV, and sleep staging on a coin cell, with a week of memory headroom.

  • Clinical & pharma research

    Large ambulatory studies where raw-waveform retention is required but storage costs scale with the cohort.

  • Industrial / animal health

    Livestock, equine, and telemetry applications operating over LoRa, sub-GHz, or NB-IoT links.

  • OEMs building on Ambiq silicon

    Drop-in INT8 encoder + C headers for Apollo-class MCUs, with server-side decoders and bring-up sample data included.

Headline benefits

  • :material-compress:{ .lg .middle } 2× – 64×

    Compression operating points, one codec

  • INT8

    Quantized on-device encoder

  • ≥ 0.96 cosine

    Waveform fidelity through aggressive ECG/PPG operating points

  • < 1 bpm

    PPG HR error across the published RVQ CR ladder

Design principle

The compression is applied once, at the sensor. Every downstream system — flash, radio, gateway, cloud ingestion, ML training — inherits the savings automatically.

Benefit 1 — On‑device memory

compressionKIT lets devices keep much more continuous waveform in the same MCU flash partition or PSRAM region.

PPG — continuous recording capacity

Frame = 5 s @ 64 Hz, 16‑bit. RVQ bit budgets: 2560 / 1280 / 640 / 320 / 160 bits per frame.

Storage Raw 2× 4× 8× 16× 32×
128 KB 1.7 h 3.4 h 6.8 h 13.7 h 27.3 h ~2.3 days
1 MB 13.7 h 1.1 days 2.3 days 4.6 days 9.1 days ~18 days
8 MB PSRAM 4.6 days 9.1 days 18 days 36 days 73 days ~5 months

ECG — continuous recording capacity

Frame = 2 s @ 256 Hz, 16‑bit. RVQ bit budgets: 4096 / 2048 / 1024 / 512 / 256 bits per frame.

Storage Raw 2× 4× 8× 16× 32×
128 KB 0.3 h 0.6 h 1.1 h 2.3 h 4.6 h 9.1 h
1 MB 2.1 h 4.3 h 8.5 h 17 h 34 h 2.8 days
8 MB PSRAM 17.1 h 1.4 days 2.8 days 5.7 days 11 days 23 days

What this unlocks

At 32× PPG compression, a single 1 MB flash partition holds two weeks of continuous PPG — enough for a full RPM study without a single BLE sync.

Benefit 2 — Radio, bandwidth & energy

Radio airtime is the dominant battery cost in most wearable designs. compressionKIT shrinks the per‑second payload directly.

Signal Raw 2× 4× 8× 16× 32×
PPG (64 Hz) 1024 bps 512 bps 256 bps 128 bps 64 bps 32 bps
ECG (256 Hz) 4096 bps 2048 bps 1024 bps 512 bps 256 bps 128 bps
3‑lead ECG 12.3 kbps 6.1 kbps 3.1 kbps 1.5 kbps 768 bps 384 bps

At aggressive ratios this enables:

  • BLE‑advertising‑only designs — a single advertisement or connection event per minute carries the full waveform
  • Sub‑GHz / LoRa / NB‑IoT telemetry for continuous cardiac data
  • Longer connection intervals → fewer wake‑ups → months on a CR2032

Benefit 3 — Cloud storage & ingestion

Per‑patient, per‑year waveform retention — and what that scales to at fleet level:

Deployment Raw @ 8× @ 32×
1 patient, PPG 24/7 4.0 GB 510 MB 130 MB
1 patient, ECG 24/7 16.1 GB 2.0 GB 520 MB
100k patient fleet, ECG 24/7 1.6 PB 200 TB 52 TB

→ ~30× lower S3 / object‑store cost, faster ETL, and — critically — faster model training over the full fleet because the input pipeline is no longer bandwidth‑bound.

Benefit 4 — Fidelity that survives downstream algorithms

Unlike naive downsampling or bit-depth truncation, compressionKIT evaluates codecs against waveform and physiology-aware scorecards. The published RVQ bundles are the easiest packages to load today; SPIHT and hybrid lanes provide standardized comparison points for clean, noisy, and artifact-heavy regimes.

  • PPG and ECG RVQ bundles include measured PRD, MSE, cosine, and modality-specific physiological metrics.
  • Noise-aware scorecards separate clean-frame faithfulness from recoverable signal quality under empirical noise and artifacts.
  • SPIHT is a strong clean-signal baseline; hybrid lanes are important when wearable noise or contact artifacts dominate.
  • ECG and PPG model pages now include the current crossover heatmaps and links to CR-vs-fidelity tables.

→ The practical question becomes which operating point preserves the observables your product needs, not which single aggregate metric is smallest.

Choosing an operating point

Every product has a different tradeoff. These are the default recommendations for each operating point:

Goal PPG ECG
Near‑lossless archival 2× 2× – 4×
Clinical‑grade HR / HRV / rhythm 4× – 8× 4× – 8×
Wellness / ambulatory monitoring 8× – 16× 8× – 16×
Event logging / screening / triage 16× – 32× 16× – 32×
Ultra‑low bandwidth telemetry 32× 32× – 64×

All PPG (2x-32x) and ECG (2x-64x) RVQ operating points ship as published bundles, with DSP/hybrid comparison lanes registered for local reproduction. See the PPG Model Zoo and ECG Model Zoo for measured metrics and noise/artifact tradeoffs at each ratio.

Getting started

  1. Try — load a published HuggingFace bundle and round-trip a synthetic frame (Getting Started)
  2. Compare — inspect RVQ, SPIHT, and hybrid tradeoffs on the model pages
  3. Evaluate — run the codec on your own waveform and score the observables you care about
  4. Deploy — validate the deploy/ package before integrating encoder/codebook artifacts into firmware

See the Deployment Guide and CLI reference for the full workflow.