β-Decode: Attention-based Decoding Temporal Artifacts via Unsupervised β-Variational Autoencoder

Physiological sensing modalities, such as Electroencephalography (EEG), Galvanic Skin Response (GSR), and Photoplethysmography (PPG), provide detailed representations of cognitive and physiological states, proving invaluable for applications in human-computer interaction and digital health. However, these time-series signals are frequently affected by stationary and non-stationary noise, temporal fluctuations, and user-specific physiological variations, compromising signal integrity. To address these challenges, we propose β-Decode, a generative, unsupervised denoising framework tailored for multi-modal time series data. β-Decode achieves two main objectives: (i) learning global and local temporal dependencies within time-series representations to enhance denoising and (ii) handling unseen temporal noise variations. β-Decode leverages a β-variational autoencoder (β-VAE) combined with an attention mechanism, capturing the data distribution via latent representations. Additionally, we introduce a modal-specific noise-coupling strategy (NCS) to simulate diverse noise patterns, enhancing β-Decode’s adaptability across datasets. We evaluate β-Decode on two public uni-modal datasets and an in-house multi-modal dataset, demonstrating that it consistently outperforms five state-of-the-art denoising algorithms, quantifying this superiority across three denoising metrics by a significant margin (e.g., Pearson Correlation Coefficients of 0.73 ± (0.0102) between original and denoised signals).