Presentation title
Energy-efficient PPG-based Heart-Rate MonitoringAuthors
Matteo Risso, Alessio Burrello, Daniele Jahier Pagliari, Simone Benatti, Luca Benini, Massimo Poncino and Enrico MaciiInstitution(s)
Politecnico di Torino, University of BolognaPresentation type
Presentation of a research group from one or more scientific institutionsAbstract
Wrist-worn devices equipped with sensors, such as wristbands and smartwatches, enable a comfortable monitoring of vital signs, hence they are becoming increasingly popular in personalized health care and medical IoT applications. Heart rate (HR) is one of the most critical indexes to monitor, both for activity tracking and for clinical purposes. First generation HR-monitoring devices were based on a simple 1-3 leads ECG, connected through a chest strip, which, however, is uncomfortable or even impossible to wear in certain conditions. Instead, thanks to the optimization and miniaturization of photoplethysmogram (PPG) sensors, it has become possible to integrate HR monitoring in smaller, less invasive and cheaper devices. A PPG sensor consists of one or more LEDs that continuously emit light to the skin and a photodiode that measures variations of light intensity caused by blood flow, which depends on the heart rate. A major limitation of PPG based HR estimation is represented by motion artifacts (MA) caused by variations of sensor pressure on the skin or ambient light leaking in the gap between the photodiode and the wrist.
Recent state-of-the-art algorithms combine PPG and inertial signals to mitigate the effect of MA. However, these approaches suffer from limited generality. Moreover, their deployment on MCU-based edge nodes has not been investigated. In this work, we tackle both the aforementioned problems by proposing the use of hardware-friendly Temporal Convolutional Networks (TCN) for PPG-based heart estimation. Starting from a single ``seed'' TCN, called TEMPONet which shows impressive results in other bio-signals analysis tasks, we leverage an automatic Neural Architecture Search (NAS) approach, by means of the novel MorphNet algorithm, to derive a rich family of models, that enable us to form two Pareto frontiers, in the Accuracy vs Complexity and Accuracy vs Network Size. From the Pareto curves designers can select a model based on the available computing resources.
All the experiments are build on top of the popular PPGDalia dataset, which is the largest publicly available collection of raw acceleration and PPG data for PPG-based HR-estimation, acquired during daily-life activities and not in a controlled scenario, such as a laboratory.
Moreover, we add to our resulting networks a simple post-processing algorithm, in charge of removing predictions that are not compatible with human physiology. Specifically, we impose a limit on the maximum relative HR variation over time.
The best performing model, BestMAE, achieves a MAE of 5.30 BPM, and includes ~232k trainable parameters. Coupling this TCN with simple post-processing and fine-tuning steps, we further reduce the MAE to 3.84 BPM, outperforming the current (more complex) state-of-the-art algorithms. At the other extreme, the smallest model in TimePPG, BestSize, uses only 5k parameters while still reaching an acceptable MAE of 6.29. Finally, as a compromise between the former two, we analyze BestMCU, i.e. the largest network that fits the memory of a popular MCU by STM, the STM32L476, which achieves a MAE of 5.64 with 41.7k parameters.
When deployed on the MCU, BestMCU consumes 5.17 mJ per inference, with a latency of 427 ms. BestSize reduces both metrics by 25x, reducing energy to 0.21 and latency to 17.1 ms.
Additional material
- Presentation slides: [pdf]