BLiSS: A Beat-Local, Signal-Quality-Informed Kalman Smoother for ECG Denoising under Non-Stationary Muscle Artifact, with Ectopic-Beat Reconstruction Assessment

Abstract ECG denoising under non-stationary muscle artifact (MA) must suppress noise while preserving ectopic morphology. We propose BLiSS, a beat-local, signal- quality-informed state-space smoothing algorithm. Following fixed wavelet- Wiener filtering and R-peak detection, raw-beat quality sets the measurement variance of a constant-velocity Kalman filter and Rauch–Tung–Striebel smoother, without neighbouring-beat morphology priors. BLiSS was compared with three Kalman-based and two EMG-focused algorithms, including adaptive wavelet- Wiener filtering with adaptive moving-average filtering (AWWF + AMAF), on non-paced MIT-BIH Arrhythmia recordings under constant-level, step, burst, and alternating MA. Targeted event-overlapping bursts assessed premature ventricular contraction (PVC) and supraventricular ectopic beat (SVEB) recon- struction. BLiSS achieved the largest average SNR gain among the evaluated methods for each contamination profile (4.44–5.38 dB), with the largest margins over AWWF + AMAF under step and burst MA. It also achieved the low- est mean PVC QRS-region normalized root-mean-square error (NRMSE; 0.217), PVC amplitude mean absolute error (0.194 mV), and SVEB QRS-region NRMSE (0.241). AWWF + AMAF achieved lower SVEB whole-beat NRMSE (0.501 versus 0.596) and lower P-peak timing error in a small atrial subgroup. Abla- tions supported beat-specific measurement weighting and backward smoothing; a preceding-beat prior increased PVC error. BLiSS therefore achieved the high- est mean denoising gain and the lowest mean ectopic QRS-region errors, while whole-beat and P-wave preservation remained limitations.