breznak / nupic.audio Goto Github PK

This is the first step of htm-community#29 which is our work for the
https://physionet.org/challenge/2016/ challenge!

The code is in https://github.com/htm-community/nupic.audio/tree/master/CinC2016Challenge_PCG

We need to preprocess:

• whole dataset from audio to a vector using matlabs wav2vec
• ... ditto using WaveEncoder (implemented)
• reduced data using FHS (see https://physionet.org/challenge/2016/)
  • faster version
  • slower version (state of the art)
  • FIRST: browse the Physionet formums for optimal params?
  • FIRST: see the forums if someone already published the preprocessing? (everyone will have to do it, just takes time; we may ask and possibly publish our preprocessed data)

The work will be done mostly in matlab in officially provided scripts + with our implemented python code.

Results will be in

• mat files
• and exported to CSV
• distinguised to Normal/Abnormal (I made a script for that from annotation)

This is a WIP, just brainstorming:

• I'm inclined to do a bag-of models (voting) classifier.
• anomaly score model on normal only
  • full data
  • FHS only
  • FHS specialized in single components
• model on Abnormal only
• models with different params
• model trained on our external ECG, ...
• spatial seq. classification (not anomaly on time-series, Union? model in nupic.research?)
• ask for other help on HTM forums, ...