An Edge-Cloud Collaborative ECG-Assisted Diagnostic System Leveraging Cross-Lead Knowledge Distillation and Large...

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Cardiology

Background:

• Cardiovascular diseases (CVDs) represent a significant global health challenge, necessitating efficient and accessible diagnostic tools.
• Real-time electrocardiogram (ECG) monitoring is crucial for timely CVD detection, but current wearable devices face limitations in diagnostic performance due to hardware constraints and spatial information loss.
• Existing lightweight models for ECG analysis often lack interpretable insights beyond basic classification.

Purpose of the Study:

• To develop an edge-cloud collaborative method for ECG-assisted analysis that overcomes the limitations of current wearable monitoring systems.
• To enhance diagnostic accuracy and interpretability of ECG analysis on resource-constrained devices.
• To enable real-time screening and cloud-side report generation for improved cardiovascular health management.

Main Methods:

• An algorithmic approach employing cross-lead knowledge distillation from a 12-lead InceptionTime-Transformer teacher model to an ultra-lightweight single-lead student model using a hybrid loss function.
• A system-level, three-layer architecture integrating edge-side real-time screening with cloud-side report generation.
• Utilized a LoRA-fine-tuned Qwen3-8B large language model for cloud-side analysis and report generation.

Main Results:

• The single-lead student model achieved 92.8% of the teacher's Macro-F1 score and 94.7% of its AUC-ROC after significant parameter compression (123.7×) and lead reduction (12-to-1).
• The quantized TFLite model (INT8) is compact (20.8 KB) and demonstrated low resource utilization (63.0 KB SRAM, 11.6 ms latency) in Cortex-M4 simulations, indicating potential for on-device deployment.
• The proposed method effectively combines lightweight model distillation with large language models for comprehensive ECG analysis.

Conclusions:

• The edge-cloud collaborative ECG analysis method demonstrates high diagnostic performance and efficiency, suitable for resource-constrained devices.
• The study successfully transfers knowledge from complex multi-lead ECG models to lightweight single-lead models, preserving diagnostic accuracy.
• Further validation on physical hardware is required to assess real-world performance, including power consumption, BLE latency, and motion artifact resilience.