---
license: other
library_name: compressionkit
pipeline_tag: other
tags:
  - compressionkit
  - signal-compression
  - ecg
  - spiht
  - wavelet
  - edge-ai
  - hybrid
  - denoising
---

# compressionkit-ecg-hybrid-2x

A **ECG** signal compression codec: a learned wavelet-gain denoiser (trained weights) followed by wavelet + SPIHT + arithmetic coding. The deployable artifact is the denoiser weights, the bitstream contract, and a portable C99 reference implementation for the SPIHT stage.

## Operating point

| Field | Value |
|-------|-------|
| Modality | ECG |
| Sample rate | 256 Hz |
| Frame size | 512 samples |
| Target CR | 2x |
| Wavelet | `bior4.4` |
| DWT levels | 6 |
| Bit budget | 4096 bits/frame |
| Entropy coder | arithmetic coding |

## Quality metrics

### Fidelity & Robustness

Both fidelity yardsticks are reported so the codec is judged fairly: **faithfulness** is PRD vs the recorded (still-noisy) input, while **truth fidelity** is PRD vs clean ground truth. Lower is better.

| Metric | Value |
|--------|-------|
| Truth PRD vs clean (%) | 7.68 |
| Truth PRD at native noise (%) | 34.61 |
| Faithful PRD vs input (%) | 8.97 |
| PRD degradation slope (PRD%/dB) | 2.42 |
| PRD at 0 dB SNR (%) | 37.88 |
| PRD at -6 dB SNR (%) | 54.38 |
| Pure-noise imprint autocorr | 0.1242 |

### Time Domain

_PRD here is faithfulness (vs the recorded input); see **Fidelity & Robustness** above for the clean-truth and noise-regime view._

| Metric | Mean | Median | P90 |
|--------|------|--------|-----|
| PRD vs input — faithfulness (%) | 8.9669 | 7.3373 | 14.5582 |
| RMSE | 0.0897 | 0.0734 | 0.1456 |
| Cosine Similarity | 0.9948 | 0.9979 | 0.9995 |

### Spectral

- **Band Total Relative Error (median):** 0.1385

### Bitrate


## Python quickstart

```python
from compressionkit.runtime import load_codec

codec = load_codec("Ambiq/compressionkit-ecg-hybrid-2x")
enc = codec.compress(frame)   # frame: (frame_size,) float32
recon = codec.decompress(enc)
```

## C quickstart

The SPIHT stage ships a portable C99 reference; the denoiser (`denoiser_gain_model.keras`) currently only has a Python/TFLite runtime path. Run the denoiser stage first (Python, or a converted TFLite Micro model) and feed its output into the C SPIHT encoder below.

```c
#include "spiht_app_config.h"

float frame[APP_SPIHT_FRAME_SIZE];
uint8_t bitstream[APP_SPIHT_MAX_BYTES];
/* ... fill frame from sensor (post-denoise, if hybrid) ... */
size_t nbits = spiht_encode_frame(&enc, bitstream, APP_SPIHT_MAX_BITS);
```

## Files

| File | Description |
|------|-------------|
| `config.json` | Deploy manifest (`family: "hybrid"`) |
| `spiht_config.json` | Codec parameters (language-neutral) |
| `sample_stimulus.npz` | Synthetic test frames |
| `reference_vectors.npz` | Reference encode/decode vectors |
| `c_sources/spiht.[ch]` | Portable C99 reference |
| `c_sources/spiht_app_config.h` | Codec-specific defines |
| `model_card.json` | Provenance metadata |
| `scorecard.json` | Frozen evaluation summary |
| `denoiser_gain_model.keras` | Learned wavelet-gain denoiser (trained weights) |
| `hybrid_manifest.json` | Pipeline stage order (denoise → SPIHT) |

## Dataset & license

Evaluation data: PTB-XL (CC BY 4.0). Sample data may include excerpts under the original license terms.

Codec source released under the **OTHER** license.

## Citation

```bibtex
@software{compressionkit,
  author = {Ambiq AI},
  title = {compressionKIT: Signal Compression for Edge AI},
  url = {https://github.com/AmbiqAI/compressionkit}
}
```