Exploring the feasibility of real-time on-device ECG biometric classification using quantized neural networks

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Wearable Technology

Background:

• Continuous personalized healthcare monitoring is crucial.
• Electrocardiogram (ECG) signals offer unique biometric patterns.
• Current systems often rely on cloud connectivity, raising privacy and efficiency concerns.

Purpose of the Study:

• To develop and evaluate a proof-of-concept embedded deep learning system for real-time ECG biometric classification on wearable Holter devices.
• To reduce reliance on continuous cloud connectivity for ECG analysis.
• To enhance data privacy and reduce power consumption in wearable health monitoring.

Main Methods:

• A quantized convolutional neural network (CNN) was deployed on an STM32H7 microcontroller.
• An initial signal quality assessment stage was incorporated for robust processing.
• The system was evaluated on the PTB Diagnostic ECG Database using subject-specific training.

Main Results:

• The embedded system achieved an F1 score of 94.51% and 94.68% classification accuracy on 5-second ECG segments.
• Average inference time was 1.35 seconds, enabling real-time operation on resource-constrained hardware.
• On-device inference demonstrated improved data privacy and reduced data transmission.

Conclusions:

• The embedded implementation proves the feasibility of integrating lightweight ECG biometrics into wearable systems.
• This approach enhances privacy, reduces power consumption, and minimizes data transmission.
• Potential for future extensions includes personalized healthcare monitoring and early anomaly detection.