Robust and Energy-Efficient PPG-Based Heart-Rate Monitoring

Abstract

A wrist-worn PPG sensor coupled with a lightweight algorithm can run on a MCU to enable non-invasive and comfortable monitoring, but ensuring robust PPG-based heart-rate monitoring in the presence of motion artifacts is still an open challenge. Recent state-of-the-art algorithms combine PPG and inertial signals to mitigate the effect of motion artifacts. 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, we leverage an automatic Neural Architecture Search (NAS) approach to derive a rich family of models. Among them, we obtain a TCN that outperforms the previous state-of-theart on the largest PPG dataset available (PPGDalia), achieving a Mean Absolute Error (MAE) of just 3.84 Beats Per Minute (BPM). Furthermore, we tested also a set of smaller yet still accurate (MAE of 5.64 -6.29 BPM) networks that can be deployed on a commercial MCU (STM32L4) which require as few as 5k parameters and reach a latency of 17.1 ms consuming just 0.21 mJ per inference.

Q-PPG: Energy-Efficient PPG-Based Heart Rate Monitoring on Wearable Devices

“…The other curves reported in Figure 6 show the results of the isolated application of the two NAS algorithms. The blue points refer to the application of MorphNet (MN) to TEMPONet with hand-tuned dilation, and correspond to the results of [19]. Orange points, instead, correspond to the application of PIT alone, using the TEMPONet variant with d = 1 as seed.…”

“…The most accurate model obtained by our automatic design space exploration, before quantization, achieves a MAE of just 4.36 BPM while requiring around 269k parameters and 17.5M OPs. On the other hand, by only using MN, as presented in [19], the best MAE obtained was 4.88 BPM, with similar number of parameters (230k) and operations (12M). Noteworthy, increasing the number of parameters from 3.5k up to 30k leads to improving the MAE from 6.5 BPM to 4.55 BPM.…”

“…The first approach works acceptably well for int8 data. For instance, in [19] we showed that the MAE degradation for HR monitoring when moving from a single-precision floating point format (fp32) to int8 was in the 1.26-1.44 BPM range. However, using QAT leads to the recovery of most of the performance loss for int8 precision, and to a limited error increase also for sub-byte precision.…”

“…In this paper, which extends [19], we propose the first systematic flow to optimize DL models for PPG-based HR tracking. We focus in particular on Temporal Convolutional Networks (TCNs), a family of DL models that are both HWfriendly and accurate for time-series processing.…”

“…All TCNs are automatically derived from a single seed architecture [20]. With respect to [19], which only optimized the number of feature maps in each TCN layer, in this work, we extend the search also to consider the dilation parameter of convolutional layers, which effectively reduces the model complexity with a limited impact on accuracy.…”