Using as few pixels from the video to do prediction (which channels? Which part of the image? e.g. using gradient to find if phone is only partly on, and then take the part where it is on the best)
[ ] noise reduction
making an as clean as possible PPG graph
e.g. w/ kalman filters
[ ] feature engineering
finding human-understandable features from the PPG that can be used in knowledge based methods
E.g. with FFT feature engineering, then feature selection, then human explainability
[ ] black-box prediction: compute-heavy
Create best possible black-box prediction method
e.g. 1d resnet
[ ] black-box prediction: compute-light
Create best possible black-box prediction method that can be run on mobile