Deep learning based heart rate estimation from PPG data with motion artifacts

Summary: Fitness/health trackers use optical sensors called PPG sensors to measure the heart rate and other health metrics of the user. For a smartwatch, the PPG sensor is located on the back making contact with the user's skin. When resting or sleeping, the sensor-skin contact is static resulting in accurate heart rate measurements. During fitness activities the sensor-skin contact changes adding what's called motion artifacts to the PPG sensor measurements. This repo implements an end-end deep learning method to estimate the heart rate accurately during dynamic scenarios. The architecture used here is an implementation of the following paper:

• Dataset: Troika
• Paper: CorNET: https://ieeexplore.ieee.org/document/8607019

Todos:

• Dataset management (added to repo for now, I know this isn't proper)
• Set up database utils code for managing the library
• Set up model and optimizer code
• Pytorch conversion -- experiences in a blog? (include mps?)
• Fix the data leakage issue and compare to paper.
• Collect all tunable parameters in a top level config file, gpu, etc
• Requirements list or dockerization
• Pytorch implementation
• Readme file update
• Get feedback and enroll e few collaborators, post
• MLX conversion
• Inference speed analysis and comparison when sped up with openvino and ONNX
• TF to PT blog and MPS vs CPU

References:

• https://github.com/KJStrand/Pulse_Rate_Estimation