Abstract
Continuous health monitoring via wearable sensors is severely bottlenecked by the energy cost of wireless data transmission. This paper introduces a Compressive Sensing (CS) framework that utilizes Online Dictionary Learning (ODL) on a remote server to create personalized signal bases, allowing wearable sensors to compress ECG data by 10x without losing diagnostic quality. By offloading complex pre-processing and learning to the server, the framework drastically reduces the data payload and extends sensor battery life.
Introduction
In Wireless Body Area Networks (WBANs), the energy required for wireless data transmission is significantly higher—often by two orders of magnitude—than the energy used for sensing or analog-to-digital conversion. To extend the operational life of wearable devices, efficient data compression is essential. While Compressive Sensing (CS) is a common solution, traditional frameworks rely on generic mathematical bases (e.g., Wavelets or DCT) that do not account for individual physiological variations. This lack of personalization limits the achievable Compression Ratio (CR). Furthermore, many existing methods require the low-power sensor to perform computationally intensive tasks like heartbeat detection, further draining battery resources.
Proposed Framework
The authors propose an asymmetric, three-tier architecture designed to shift the computational burden away from the wearable sensor.
- Sensor Node (Data Acquisition): Performs simple data segmentation and random encoding using a Bernoulli matrix to compress the signal with minimal CPU usage.
- Server Node (Online Dictionary Learning): Executes Online Dictionary Learning (ODL) to develop a personalized dictionary of "atoms" specific to the patient's ECG patterns. The server also handles signal filtering (notch and band-pass). ODL is utilized for its scalability and ability to process samples sequentially.
- Mobile Node (Reconstruction): A smartphone or similar device uses the Basis Pursuit (BP) algorithm and the personalized dictionary to reconstruct the high-quality ECG signal.
Results and Discussion
The framework was evaluated using the MIT-BIH Arrhythmia Database, consisting of over 600,000 samples. Results demonstrate that personalization through ODL unlocks extreme signal sparsity, allowing for significantly higher compression.
The proposed framework achieved a 10x Compression Ratio while maintaining an average Percent Root-mean-square Difference (PRD) of 9%, which is the standard threshold for clinical diagnostic quality. This represents a 2-4x improvement over state-of-the-art CS frameworks that utilize pre-determined bases. Specifically, compared to a joint DCT-DWT basis, the personalized ODL approach allows for a 6.5x greater reduction in sample size at the same quality level, leading to proportional energy savings in the sensor's radio transmitter.
Conclusion
By moving the heavy lifting of dictionary learning and pre-processing to a remote server, this framework enables highly efficient, long-term ECG monitoring on low-power wearable devices. Future work will focus on incorporating a "sub-basis" system to better handle sudden abnormal ECG signals that may deviate from a patient's learned baseline.
Summary updated