08 / Signal processing research · 2026
ECG denoising with a verified noise protocol
The target SNR is measured, and tuning records stay separate from evaluation records.
The audit found that the noise coefficient was divided instead of multiplied. The formula, tests and experimental split were corrected before the result was returned to the portfolio.
Evidence register
6/6
tests
Including observed-SNR checks
R-peaks
segmentation
Physiology-aware helper
0
record overlap
Tuning and evaluation records are disjoint
01 / Problem
An inverted noise scaling formula made the announced 5 dB experiment unreliable and same-data selection weakened the evidence.
02 / Approach
Scale noise by the derived amplitude factor, assert the observed SNR and isolate records used for selection from final evaluation.
03 / Outcome
The numerical contract is testable and the experiment no longer depends on an unverified SNR claim.
System trace
How the evidence is produced.
Audited the mathematical implementation, repaired the protocol and added reproducible safeguards and CI.
- 01Clean ECG → record split
- 02Noise model → target SNR
- 03Denoiser → reconstructed signal
- 04Held-out records → metrics
Validation scope
Six tests pass locally and in CI, covering noise scaling and core data boundaries.
Known limitation
The repaired protocol is not clinical validation and does not yet establish cross-dataset or cross-patient generalization.
What is inspectable
- Critical formula corrected.
- Observed rather than assumed SNR.
- Selection and evaluation data separated.
Next proof to add
- 01Evaluate morphology-aware metrics.
- 02Benchmark multiple noise types.
- 03Add cross-database patient-level validation.
Main stack