Time Scale Network: An Efficient Shallow Neural Network For Time Series Data in Biomedical Applications.

Area of Science:

• Biomedical data analysis
• Deep learning for time series
• Signal processing

Background:

• Biomedical time series data often contains information across multiple time scales.
• Existing deep learning models can be computationally expensive, data-intensive, and difficult to interpret.
• These limitations hinder real-world applications, especially in resource-constrained environments or for real-time processing.

Purpose of the Study:

• To introduce a computationally efficient deep learning network for analyzing multi-scale time series data.
• To reduce network size, data requirements, and computational complexity while maintaining or improving performance.
• To enable easier interpretation and application in real-time or edge computing scenarios.

Main Methods:

• Developed a novel Time Scale Network integrating discrete wavelet transform principles with convolutional neural networks.
• Employed back-propagation for training the network to learn features across various time scales simultaneously.
• Evaluated the network on Atrial Dysfunction detection using ECG signals and seizure prediction using EEG signals.

Main Results:

• The Time Scale Network demonstrated superior accuracy per parameter and per operation compared to traditional methods.
• Achieved fast training and inference speeds, enabling efficient real-time analysis.
• Successfully detected atrial dysfunction with interpretable learned patterns and predicted seizures with 90.9% accuracy using minimal parameters (1,133).

Conclusions:

• The proposed Time Scale Network offers a computationally efficient and interpretable solution for multi-scale time series analysis.
• It significantly reduces resource demands, making it suitable for real-world biomedical applications and edge devices.
• The method is versatile and applicable to any time series data containing features at multiple time scales.