Validation Scorecard¶

A framework for validating that lossy compression preserves physiologically and clinically relevant features — not just waveform fidelity.

Core claim¶

The codec is a task-preserving, morphology-bounded, versioned lossy compression system for ECG and PPG, with explicit validation of clinically and analytically relevant features.

Evidence layers¶

1. Waveform fidelity — Time-domain and frequency-domain reconstruction error.
2. Physiological feature fidelity — HR, HRV, pulse timing, ECG morphology, PPG AC/DC, SpO2-sensitive ratios.
3. Downstream task preservation — Same algorithms run on raw and reconstructed signals produce non-inferior outputs.
4. Operational safety — Boundary handling, fallback behavior, bitrate adaptation, versioning, and decoder compatibility.

Measurement philosophy¶

Agreement vs accuracy¶

Most metrics in this scorecard measure agreement between raw and compressed signals (same algorithm applied to both). This is distinct from task accuracy (compressed vs ground truth). Agreement tells you "did compression break anything?" but is bounded by the algorithm's own repeatability.

Algorithm repeatability baseline¶

Before attributing a metric delta to the codec, establish the algorithm's inherent jitter:

• Run the same metric on the raw signal with ±1 LSB noise injection (10 repetitions)
• The standard deviation of those measurements is the measurement floor
• Compression error within the measurement floor is indistinguishable from algorithm noise

Noise preservation vs noise removal¶

A codec that denoises (e.g., learned RVQ) may show higher PRD against a noisy ground truth but better downstream task performance. Conversely, a signal-agnostic codec (e.g., SPIHT) preserves noise faithfully — lower PRD but potentially worse clinical metrics at high compression.

The scorecard must report both:

• PRD / PRDN (noise-adjusted PRD)
• Downstream metric agreement (HR, HRV, morphology)

A codec that "hides" bad signal quality (low SQI raw → high SQI reconstructed) is dangerous. The scorecard explicitly checks for this.

Implementation tiers¶

ECG validation¶

HR / HRV preservation¶

Morphology preservation¶

Downstream task preservation¶

PPG validation¶

HR / Pulse-rate preservation¶

Morphology preservation¶

SpO2-sensitive preservation¶

Downstream task preservation¶

Cross-cutting validation¶

Stitching and streaming behavior¶

Noise and artifact behavior¶

Adaptive bitrate and fallback¶

Versioning and compatibility¶

Signal condition buckets¶

Metrics should be reported per-bucket, not only in aggregate.

ECG buckets¶

• Clean sinus rhythm
• Low-amplitude ECG
• Baseline wander
• Muscle artifact
• Electrode motion
• Premature atrial contractions (PAC)
• Premature ventricular contractions (PVC)
• Atrial fibrillation
• Bradycardia / tachycardia
• QRS widening
• ST elevation / depression
• T-wave abnormality
• Noisy but clinically interpretable

PPG buckets¶

• Clean resting PPG
• Low perfusion
• High motion
• Loose sensor fit
• Ambient light contamination
• Exercise transition
• Sleep
• Cold / vasoconstriction
• Arrhythmia-affected pulse intervals
• SpO2 desaturation (if available)
• Noisy but physiologically interpretable

Positioning guidance¶

Avoid:

"The model learns what is important."

Prefer:

"The codec is constrained and validated to preserve specific physiological observables — timing, morphology, spectral content, baseline behavior, AC/DC structure, and downstream algorithm outputs. When preservation confidence is low, the system can increase bitrate or transmit fallback residuals."

Pass/fail summary (per modality and profile)¶

For each codec version, report:

1. Average performance
2. p95 performance
3. Worst-case bounded error
4. Performance by signal-quality bucket
5. Performance by pathology/artifact bucket
6. Boundary-specific performance
7. Downstream task agreement
8. Fallback/escalation rate
9. Version compatibility status

A codec version is releasable for a profile only if it passes all required preservation bounds for that profile.