Tinybionet: A lightweight time-frequency network for biomedical signal classification on edge devices

Area of Science:

• Biomedical Signal Processing
• Edge Computing
• Deep Learning

Background:

• Edge devices face challenges in real-time biomedical time-series analysis due to signal non-stationarity, noise, and computational limits.
• Standard deep learning models are often too computationally intensive for embedded systems.

Purpose of the Study:

• To propose TinyBioNet, a compact convolutional neural network for efficient affective state classification on resource-constrained edge devices.
• To enable real-time analysis of biomedical signals directly on embedded hardware.

Main Methods:

• Developed TinyBioNet, a compact CNN with an embedded fixed-basis time-frequency transformation using parallel 1D convolutional layers.
• Utilized depthwise convolutions and residual connections for efficiency.
• Applied aggressive low-bit quantization for embedded optimization.

Main Results:

• TinyBioNet achieves state-of-the-art performance with only 5.6k parameters.
• Demonstrated robust generalization across three public biomedical signal datasets.
• Achieved up to 98.89% accuracy for PPG and 99.38% for ACC signals with negligible performance loss under 4-bit quantization.

Conclusions:

• TinyBioNet offers an efficient solution for real-time affective state classification on edge devices.
• The integrated time-frequency transformation and compact architecture enable high performance with minimal resources.
• Low-bit quantization maintains accuracy, making the model suitable for practical embedded deployment.