PULSE (Heart Rate Detection)
PULSE goal was to develop a deep learning model that takes as input PPG
(Photoplethysmogram) signals originated by wearable devices (e.g., smartwatch) and fuse them
with triaxial accelerometer signals to estimate the user’s heart rate. The developed model
outperforms state-of-the-art algorithms using Mean Absolute Error (MAE) as metric, while it is
very light weight and is able to run in real-time on low-power edge devices.
What is the challenge that is being addressed?
Time Series Analysis provides details on a signal such that condition information can be estimated. It has important application within Patient Monitoring for human Heart Rate Detection, providing heart rate information compensated for movement. In particular, a PPG (photoplethysmography) is a widely used sensor to estimate heart rate and is placed on wearable devices (smartwatches). A PPG sensor observes the capillaries in the wrist which fill with blood when the heart ventricles contract. The light emitted by the PPG sensor is absorbed by red blood cells in these capillaries and a photodetector will see the drop in reflected light. When the blood returns to the heart, fewer red blood cells in the wrist absorb the light and the photodetector sees an increase in reflected light. The period of this oscillating waveform provides the pulse rate. When the PPG is combined with a 3-axis accelerometer in a wearable can provide robust and meaningful under real world conditions information for patients in all activity levels from rest to extreme levels of movement.
What is the AI solution the project plans to implement?
PULSE is based on a deep learning model that takes as input PPG signals originated by wearable devices (e.g., smartwatch) and fuse them with triaxial accelerometer signals to estimate the user’s heart rate. The model relies on several dilated 1D convolutions followed by a multi-head attention layer that is able to extract correlation between the PPG and accelerometers signals, providing also attentional maps that can be used as a form of explanations. The developed model outperforms state of-the-art algorithms using Mean Absolute Error (MAE) as metric, while it is very light weight and is able to run in real-time on low-power edge devices.
How will BonsAPPs support you in implementing this solution?
BonsAPPs really helped us to develop a MLOps-wise solution for wearable-based heart rate estimation. The use of BonsAPPs modular services made it possible to train, compress, optimise, benchmark, and deploy our models on selected edge devices effortlessly. Moreover, during the lifetime of project, we received technical and business support coordinated by experienced technical and business mentors respectively, which was very helpful. Finally, Bonseyes Distributed AI Marketplace will help us monetize our solution by delivering this real-world solution to the industry.