search

nealgravindra / wearables

MIT License
2 stars 1 forks source link

compare baseline measurement to later ones to see if and why predictability changes and track down CNN w/high-perf #9

Open nealgravindra opened 3 years ago

nealgravindra commented 3 years ago

Based on this analysis: https://pubmed.ncbi.nlm.nih.gov/31018734/ it will be important to correlate changes in each pt's ts with progression of pregnancy

nealgravindra commented 3 years ago

along these lines, another type of post-hoc inference can be train model, then compare diff_maps of saliency and comparitive importance of onset of sleep, first activity time, etc.

nealgravindra commented 3 years ago

or build a sort of inceptiontime track for embedding just the ts, then a feature engineering/or extracte features track (e.g., https://academic.oup.com/annweh/article/64/4/350/5735350 extract sleep parameters)

nealgravindra commented 2 years ago

Also could modify CNN based on modifications to VGG described in model section in Uchida PLOS ONE 2021 review

nealgravindra commented 2 years ago

Try contrastive paired loss to see if enhancement with new loss can be found and whether this reduces the error above from IT with errro on paired measurements vs. contrastive loss (see: https://proceedings.neurips.cc/paper/2018/file/6bf733bb7f81e866306e9b5f012419cb-Paper.pdf)