Biomedical signal processing project based on an INTERCON paper

ECG Signal Denoising

Benchmark of DWT, PCA, and Kernel PCA on noisy cardiac signals

This project focuses on signal processing rigor. It benchmarks denoising methods on cardiac waveforms corrupted by muscle artifact, electrode motion, and white noise, then compares the resulting reconstruction quality.

View code

ECG records

Noise types

2.57

Best mean MSE

Problem

Biomedical signals are easily degraded by multiple noise sources, and denoising choices should be compared systematically rather than selected heuristically.

Approach

Loaded MIT-BIH records, added controlled noise, segmented beats, applied DWT, PCA, and Kernel PCA denoising strategies, and measured reconstruction quality through MSE across records and noise types.

Results

Across the benchmark, Kernel PCA achieved the lowest mean MSE overall at 2.57 versus 3.99 for DWT and 18.80 for PCA, with especially strong gains on electrode motion noise.

What is in the repository

Benchmarked three denoising families across several records and noise conditions.

Reproduced a paper-inspired setup on real cardiac datasets rather than toy signals.

Surfaced that the best method depends on the noise regime, which is a useful modeling nuance.

Packaged the workflow with separate benchmark and hyperparameter search scripts.

Role and scope

Signal processing pipeline implementation, benchmarking, and result synthesis

Project context

Biomedical signal processing project based on an INTERCON paper

Main stack

PythonWaveletsKernel PCANumPySciPywfdb