From Biosignals to Bedside: A Review of Real-Time Edge Machine Learning for Wearable Health Monitoring

Area of Science:

• Biomedical Engineering
• Machine Learning
• Wearable Technology

Background:

• Wearable devices collect diverse biosignals (ECG, PPG, etc.) for continuous, real-world health monitoring.
• Edge machine learning on wearables or phones enables privacy-preserving, real-time data analysis.
• Clinical adoption is hindered by noisy data, delayed labels, and evolving device ecosystems.

Purpose of the Study:

• To provide a comprehensive review of the end-to-end deployment pathway for edge machine learning on wearable biosignal data.
• To synthesize literature on sensing, data processing, model selection, and deployment engineering for on-device inference.
• To offer practical guidance and reusable templates for clinical translation and reliable system deployment.

Main Methods:

• Narrative review organizing literature by deployment stages: sensing, data processing, and on-device models.
• Comparison of classical feature-based methods with learned representations (CNN, TCN, recurrent, attention models).
• Synthesis of deployment engineering techniques (quantization, pruning, distillation) and benchmarking for runtime constraints.

Main Results:

• Detailed comparison of model families and their suitability for edge deployment based on resource constraints (latency, RAM, energy).
• Identification of key applications in cardiovascular monitoring, hemodynamics, sleep, respiration, movement, and stress.
• A proposed validation ladder and reliability toolkit for calibration, uncertainty management, drift detection, and update governance.

Conclusions:

• Successful clinical adoption requires addressing data quality, model efficiency, and robust validation.
• Deployment-oriented synthesis guides model selection based on edge capabilities and resource budgets.
• Reusable reporting templates, including edge-cost cards, facilitate transparent and reproducible deployment.