A Self-Adaptive Frequency Decomposition Approach for Denoising to Enhance Data-Driven Learning of Cyclic Time Series in Medical Signal Estimation

In recent advancements within the data-driven world, cyclic time series have become indispensable in industries such as healthcare, manufacturing, and energy. Deep learning methodologies have significantly improved predictions for cyclic time series across these domains. This study investigates the impact of various filtering techniques on the accuracy of deep learning models in medical signal estimation, specifically using photoplethysmography (PPG) signals as input. We argue that conventional filtering techniques, when applied as preprocessing to noisy signals, often distort the waveform and hinder the representation learning capabilities of deep learning models. Additionally, diverse and dynamic noise sources in clinical environments remain significant challenges to overcome. To address these challenges, we propose a self-adaptive frequency decomposition approach with regularization weighting policy. Through experimentation with different combinations of filtering methods and sequence-to-sequence deep learning models, our results demonstrate a 45% improvement in MSE over the Butterworth filter, 56% over the FIR filter, and 18% over the wavelet-based filter, confirming that self-adaptive decomposition filters offer superior performance for noisy cyclic time series in deep learning applications.