High-density biosignals capture spatial and temporal information about tissue activation patterns, enabling detailed mapping of physiological processes across anatomical regions. However, the enhanced spatial resolution from densely packed sensor arrays comes with substantial visual complexity and computational/memory costs. Built upon recently developed time frame processing, we propose a temporospatial frame-based biological series representation (BioTsF) that follows the principle that local tissue excitements drive movements over time. Specifically, the information from a multivariate series on specific sensors (spatial) and time intervals (temporal) in data frames is processed instead of the traditional data points through time, mirroring how multiple tissue movements within muscles are invoked. As a proof of concept, the maximum of the frame (MF) and average true range (ATR) features, joined with linear discriminant analysis (LDA) and neural network (NN) classification methods, are examined on a 256-channel high-density surface EMG (HD-sEMG) dataset over standard, faulty, and shifted data scenarios. Standard and shifted scenarios assess classifications in intra- and inter-session contexts, respectively, while faulty scenarios evaluate classifications with missing sensory information due to faults. Results show that the proposed BioTsF is more accurate than the state-of-the-art models by up to 7% (p-value=3e-11), especially in shifted scenarios, and it achieves significantly less memory allocation of up to 0.000125%. Furthermore, the proposed method reveals more understandable HD data visualization and post hoc decision explanations in both dynamic and maintenance tasks to gain physicians’ trust. Overall, the present work could significantly contribute to less memory storage, efficient computation, and higher robustness of high-density signal processing.

Keywords