Low Complexity Binarized 2D-CNN Classifier for Wearable Edge AI Devices.

Area of Science:

• Wearable Artificial Intelligence-of-Things (AIoT)
• Biomedical Signal Processing
• Low-Power Hardware Acceleration

Background:

• Wearable AIoT devices demand high resource and energy efficiency for practical application.
• Existing models often struggle to balance performance with computational constraints in wearable settings.

Purpose of the Study:

• To introduce a novel quantized multilayer perceptron (qMLP) model for efficient ECG signal processing.
• To enable accurate classification of ECG signals using a binary convolutional neural network (bCNN) on resource-constrained hardware.

Main Methods:

• Developed a quantized multilayer perceptron (qMLP) for transforming ECG signals into binary images.
• Integrated the qMLP with a binary convolutional neural network (bCNN) for signal classification.
• Deployed the combined model onto a low-power Field Programmable Gate Array (FPGA) fabric.

Main Results:

• Achieved a classification accuracy of 98.5%, sensitivity of 85.4%, specificity of 99.5%, precision of 93.3%, and F1-score of 89.2%.
• Reduced multiply-accumulate (MAC) operations by 5.8× compared to existing wearable CNN models.
• Demonstrated a dynamic power dissipation of only 34.9 μW.

Conclusions:

• The proposed qMLP-bCNN model offers a highly efficient solution for wearable AIoT applications.
• The model achieves state-of-the-art performance with significantly reduced computational and power requirements.
• FPGA deployment proves the feasibility of implementing advanced AI models on low-power wearable devices.