Abstract
Multiscale convolutional neural networks are widely utilized in the analysis of physiological signals to assess human health due to their large receptive field and rich feature domain. However, current scale selection for multiscale networks relies heavily on data and empirical knowledge, leading to information redundancy and reduced generalization capability, which hinders clinical application. To address these, we propose a novel paradigm of deformable information gain-guided scale (DIGS), which adaptively selects the appropriate scale for feature extraction according to the input signal. Specifically, we employ an information entropy-based hierarchical calculation of the information gain between the input and output features of each multiscale branch. This information gain determines the optimal branches for feature extraction. Additionally, we utilize 1-D deformable module to achieve overall convolutional shifts, enabling scale information to match the distribution of the dataset. Extensive experiments conducted on five datasets demonstrate that our proposed multiscale paradigm outperforms previous baseline networks in four classification tasks of physiological signals. Furthermore, integrating our structure with the best performing network on current private wearable PPG datasets improves accuracy by approximately 1%, sensitivity by 1%–3%, and reduces computational load by nearly half. These findings indicate that our new structure enhances both classification performance and computational efficiency of multiscale networks.
Funding Information
  • Key Laboratory of Sensing Technology and Biomedical Instrument of Guangdong Province
  • Shenzhen Science and Technology Plan for fundamental research (JCY20190807162003696)

This publication has 52 references indexed in Scilit: