WaveMamba-Net: Dual-frequency adaptive wavelet state-space network for real-time pulse signal classification in wearable health monitoring

<h2>Summary</h2> Continuous cardiovascular monitoring via wearable devices is critical for early disease detection, yet existing pulse signal analysis methods struggle to achieve both high accuracy and real-time performance under noisy, imbalanced conditions. We propose WaveMamba-Net, a deep learning framework integrating wavelet-based multi-scale decomposition with state-space modeling for patient status classification from photoplethysmography signals. Discrete wavelet transform decomposes pulse signals into frequency-specific components, processed by dual-view Wavelet Temporal and Spatial Sliding Enhance Modules capturing interperiod rhythm and intraperiod morphological features. Selective state-space Mamba blocks enable efficient long-range dependency modeling, while Multi-Wavelet Condition Convolution adaptively integrates decomposition levels via input-dependent kernel modulation. Evaluated on Medical Information Mart for Intensive Care III (MIMIC-III) Waveform Database and PPG-DaLiA Dataset, WaveMamba-Net achieves 91.84% accuracy and 87.52% Macro-F1 for arrhythmia classification, and 88.26% accuracy with 86.91% Macro-F1 for stress assessment, outperforming nine baselines with 12 ms inference latency suitable for edge deployment.