Efficient and Robust Heart Rate Estimation Approach for Noisy Wearable PPG Sensors Using Ideal Representation Learning
← back to projectsOverview
This research project explored robust and efficient heart rate estimation from noisy wearable photoplethysmography (PPG) signals.
Wearable PPG sensors are widely used in smartwatches and mobile devices but are highly susceptible to motion artifacts, sensor variability, and biological noise. Instead of learning to model every possible noise condition, this project introduces a new strategy: learning from mathematically modelled ideal PPG signals that represent clean cardiac dynamics.
Using a generative learning framework, the system converts noisy real-world signals into simplified representations that preserve heart-rate-related information while suppressing noise. The result is a lightweight and generalizable model suitable for real-time deployment on resource-constrained devices.
Vision
Wearable health sensing is increasingly embedded into everyday life, yet real-world noise remains a major barrier to reliable physiological monitoring.
This project aims to shift the focus from noise modelling toward ideal representation learning, where models learn the underlying physiological structure rather than dataset-specific noise patterns.
The long-term vision is to enable:
- Reliable heart rate monitoring across diverse devices and users
- Robust performance under motion and real-world activity
- Lightweight models suitable for mobile and wearable hardware
- Generalizable physiological sensing beyond heart rate estimation
Example Applications
The project supports several practical and research applications:
- Wearable Heart Rate Monitoring: Improved HR estimation during movement and daily activities.
- Mobile Health Applications: Real-time inference on smartphones without requiring GPUs.
- Cross-Device Generalization: Robust performance across different PPG sensors.
- Future Health Analytics: Foundation for estimating additional signals such as pulse transit time or blood pressure.
Evaluation
The approach was evaluated across multiple public datasets and a dedicated user study involving unseen participants and novel hardware.
Key findings include:
- Improved heart-rate estimation accuracy compared to state-of-the-art methods under noisy conditions
- Strong generalization across activities such as sitting, walking, and running
- Reduced model complexity (~12× smaller than prior approaches)
- Real-time inference demonstrated on an Android mobile device
These results highlight the practicality of ideal-representation learning for wearable sensing systems.
Open Source & Availability
The project is released as an open-source implementation:
https://github.com/aid-lab-org/Ideal-PPG-HR
Get Involved
We welcome collaborations from researchers and developers interested in wearable sensing, mobile health, and robust machine learning for physiological signals.
If you are interested in using or extending this work, please contact the authors.