In this paper, we present a novel wavelet-based approach for physiological signal analysis. To address the issues of motion artifacts, baseline drift, and low SNR noise in physiological signals, we leverage the wavelet transform to capture multi-scale time-series-frequency features of various physiological signals. We present the first large-scale pre-trained models specialized for EMG and ECG, which outperform existing methods, and build an integrated multi-modal framework integrating EEG models to effectively address the issues of low SNR, high inter-individual variability, and device mismatch, and achieve superior performance over existing methods in multi-modal tasks. The wavelet-based architecture lays the foundation for diverse physiological signal analysis, and the multi-modal design presents the next generation of physiological signal processing that will impact wearable health monitoring, clinical diagnosis, and a wide range of biomedical applications.
