MW-BLAS: Multi-Window Bayesian Look-Around Search for Robust Peak Detection in Complex Quasi-Periodic Signals

Peak detection in physiological signals is fundamental to biomedical signal processing, yet existing methods exhibit systematic performance degradation under non-stationary conditions: fixed-threshold methods suffer precision-recall drops exceeding 30% when applied to amplitude-modulated or frequency-swept signals [4]. We propose Multi-Window Baseline-corrected Local Amplitude Search (MW-BLAS), a robust peak detection algorithm employing multi-window analysis using adaptive voting-based consensus. MW-BLAS applies baseline correction using multiple window sizes matched to expected signal periodicities, then identifies peaks through weighted voting across scales. An adaptive time-weighting mechanism enables tracking of continuously varying signal characteristics. Performance was evaluated on four synthetic benchmarks: a stationary sinusoid (B1) and a multi-frequency superposition(B2). Two non-stationary signals were also tested: an amplitude-modulated signal (B3) and frequency-modulated chirp (B4). All methods were compared against find_peaks_cwt (python scipy package) and single-window BLAS. While all methods achieved perfect performance (F1 = 1.000) on stationary signals (B1), MW-BLAS significantly outperformed baseline methods on non-stationary benchmarks. Mean F1-scores across B2–B4 were 0.879 for MW-BLAS versus 0.758 for BLAS and 0.537 for find_peaks_cwt, representing relative F1-score improvements of +16.0% over BLAS and +63.7% over find_peaks_cwt. MW-BLAS achieved the highest F1-score in all non-stationary conditions: 0.822 (B2), 0.923 (B3), and 0.891 (B4). It is observed that on the chirp signal (B4), MW-BLAS improved F1-score by +175.9% over find_peaks_cwt through adaptive time-weighted voting. MW-BLAS provides robust peak detection across diverse signal conditions through multi-scale consensus voting and adaptive weighting. The method demonstrates advantages for signals with time-varying spectral characteristics, offering a practical solution for automated analysis of complex physiological recordings.