Introduction
The rapid proliferation of wearable health technologies, such as smart patches and fitness trackers, has created a significant demand for continuous, real-time analysis of physiological signals. However, these devices are constrained by a formidable "energy wall." Detecting R-peaks in an Electrocardiogram (ECG)—the most critical step for heart rate monitoring—is computationally intensive. While advanced methods like deep learning offer high accuracy, they often consume excessive power, limiting the operational lifespan of battery-powered wearables. This research addresses the challenge of implementing a robust, clinical-grade heart rate detection system that is optimized for ultra-low-power (ULP) platforms.
Methodology
The researchers optimized the gold-standard Pan-Tompkins (PT) algorithm for two distinct hardware targets: a commercial ARM Cortex-M4 (STM32) and an ultra-low-power RISC-V platform known as Vega, developed as part of the PULP project.
Hardware Implementation
The experimental setup involved a custom-built hardware chain, including electrode placement on the forearms, an acquisition board featuring the Maxim MAX30003 analog front-end, and the processing units (STM32-NUCLEO and Vega).
Algorithmic Optimizations
To achieve high efficiency without compromising accuracy, several technical optimizations were implemented:
- Fixed-Point Arithmetic: The algorithm was ported to 24-bit signed fixed-point numbers to avoid the energy overhead of Floating Point Units (FPU).
- Memory Management: Circular buffers were utilized for real-time data streaming, minimizing latency and memory movement.
- Streamlined Pipeline: The signal undergoes a multi-stage process: DC cancellation, band-pass filtering (5–12 Hz), differentiation, squaring, and Moving Window Integration (MWI) before final peak detection.
- Pruning: Redundant detection steps from the original PT design were removed based on experimental verification to reduce the total cycle count.
Experimental Setup and Datasets
The system was validated across four diverse scenarios to ensure reliability under various conditions:
- Normal Sinus Rhythm (NSR): Baseline heart activity from the MIT-BIH database.
- Atrial Flutter (AFL): Pathological conditions to test arrhythmia detection.
- High-Intensity Exercise (HIE): Testing robustness against motion artifacts and rapid heart rate changes.
- Real-Time (RT) Data: Direct acquisition using the custom hardware setup.
Key Results
The optimized system demonstrated exceptional performance across all metrics:
- Accuracy: The system achieved a consistent accuracy of over 99.5%, reaching 99.95% for normal rhythms and maintaining 99.53% even during high-intensity exercise.
- Energy Efficiency: The RISC-V Vega platform proved significantly more efficient, consuming only 0.203 mJ for a 25-second processing window—nearly 13 times less than the ARM Cortex-M4 (2.652 mJ).
- Throughput: Vega processed 61.35 samples per millisecond, more than double the throughput of the ARM-based setup.
- Operational Longevity: Based on these results, a wearable device powered by a 100 mAh battery could provide approximately 81 hours of continuous, real-time monitoring.
Conclusion
This work demonstrates that a well-optimized "classic" algorithm can outperform more complex modern approaches in the ultra-low-power domain. By leveraging hardware-software co-design on the RISC-V architecture, the researchers achieved a significant reduction in energy consumption while maintaining clinical-grade accuracy. Future work will explore the integration of lightweight machine learning models for arrhythmia classification and the use of multi-core parallelization to further enhance processing capabilities within the same power budget.
Summary updated