Comparative Performance Assessment of Stationary Wavelet Transform Configurations for ECG Denoising Using Multi-Metric Evaluation Framework

The electrocardiogram (ECG) is widely used in cardiovascular diagnosis but is frequently corrupted by noises such as muscle artefacts, baseline wander, electrode motion, and power-line interference, which can obscure clinically important features including the P wave, QRS complex, and T wave. This study presents a retrospective comparative ECG denoising framework based on the Stationary Wavelet Transform (SWT) with systematic parameter assessment across 615 scheduled configurations per evaluated case, covering five wavelet families (sym4, sym8, db4, db6, coif5), wavelet-specific decomposition levels (3–7 or 4–8) and threshold scaling factors (k = 0.5:0.1:2.8 or 0.6:0.1:3.0). The highest-scoring configuration is selected using a heuristic composite scoring function integrating output signal-to-noise ratio (SNR), percentage root-mean-square difference (PRD), Pearson correlation, and QRS timing error. Root-mean-square error (RMSE) and SNR improvement are reported separately and are not score components. The method is evaluated on twelve noise-condition cases derived from two underlying MIT-BIH recordings (records 118 and 119) in the MIT-BIH Noise Stress Test Database (NSTDB), with paired MIT-BIH Arrhythmia Database (MITDB) signals used as references. Results show SNR changes ranging from −4.17 to +8.01 dB, with larger positive changes generally observed among the lower-input-SNR cases. Sym4 and db4 wavelets were selected most often in the tested cases. Their compact support may support temporal localisation, but this observation does not establish diagnostic morphology preservation. The level 3, low-threshold setting (k = 0.50) was frequently selected under the tested ranking rule. Correlation values ranged from 0.21 to 0.97; five of twelve cases exceeded 0.90. QRS timing error was below 15 ms in seven cases and above 15 ms in five cases, providing signal-level timing observations rather than clinical validation. Denoising can degrade higher-SNR ECG signals through over-smoothing, while substantial residual distortion remains under severe noise. Overall, the framework compares candidate SWT configurations under a paired-reference protocol; it does not establish a generalized configuration or a reference-free online selector. The analysis used one fixed 10-s segment for each case, and the findings should therefore be read as controlled segment-level results.