READ ME for SIP 2020

TODO list

Analysis/Exploration/Visualization

• Find ways of displaying data categories (pairplots and histograms)

• Run any machine learning classifier (use the code from the jupyter notebook and the pyWholeBrainClassify.py)

• Investigate based on different age groups:

• Compare between domain vs no-domain (x2 per other comparison)

• Compare each age group to all the data together (total of 8 ML algorithms) Use randomForest for classification

• Needed plots

  • Plot the two features from no-domain data to see how they separate (Saathvik)
  • Plot metrics on top of t-SNE plot domain plot (including age and animal number; Saathvik)
  • Create plot to show the extent of the data we showing (how many eigenvectors, how many ages?; Anna)
  • Implement contours intstead of a bunch of scatter plots? (Use 2dhistogram and then countours, Giovanni)
  • Combine the no-domain and domain accuracy plots together (overall accuracy; Saathvik)

• New Data

  • Get new data for analysis, run through our model (Brian)
  • Run new data through model, compare to how the new data performs on the data

Presentation

Remember: We probably won't show all the data/things we try, since we have limited amount of time. We need to create a good representation even if its not everything.

Next practice presentation: Friday August 7, 2020 at 3PM

- Desi
- Sydney

• Create and manage a script for the presentation/manage the organization of the slides (Anna).
  • Add informative slide titles.
  • Periodically make the slides connect to the bigger picture.
  • Make sure all plots have axis, labels, colorbars, legends, etc.
• Make an eigenvector slide that describes how metrics were measured (Giovanni).
  • Spend a lot of time developing the spatial and temporal features.
  • Fourier analysis description.
• Add human classification of artifacts vs signal (Saathvik)
  • show confussion matrix?
  • Make it more prominant/this is our anchor

FINAL PRESENTATION: Saturday August 15, 2020

==================================================================

New list:

['temporal.autocorr', 'region.extent', 'region.majaxis', 'region.minaxis', 'mass.region', 'threshold.area', 'freq.rangesz', 'freq.maxsnr.freq', 'freq.avgsnr', 'temporal.max', 'age']

Note: I added 'temporal.autocorr' and 'region.extent' because they work well when I run the whole dataset through a random forest classifier with 18 branches and 5 max features per branch.

Age-wise list of best-performing features:

P1: ['temporal.min','temporal.autocorr','region.extent','region.majaxis','region.majmin.ratio',"region.minaxis","mass.region","threshold.area","mass.total","freq.rangesz", "freq.maxsnr.freq", "freq.avgsnr", "spatial.min","temporal.max"]

P2: ['freq.rangesz', 'freq.maxsnr.freq', 'freq.avgsnr','temporal.autocorr', 'temporal.min', 'temporal.max', 'mass.region', 'region.majaxis', 'region.minaxis', 'threshold.area', 'spatial.min','region.extent']

P3: ['mass.region', 'region.majaxis', 'region.minaxis','threshold.area', 'temporal.autocorr','spatial.min', 'temporal.min', 'region.extent','temporal.max', 'freq.maxsnr.freq']

P4: ['temporal.autocorr', "freq.rangesz",'region.extent', 'region.majaxis', 'region.minaxis', 'mass.region', 'threshold.area', 'freq.maxsnr.freq', 'freq.avgsnr', 'temporal.max']

P5: ['temporal.autocorr', "freq.rangesz",'region.extent', 'region.majaxis', 'region.minaxis', 'mass.region', 'threshold.area', 'freq.maxsnr.freq', 'freq.avgsnr', 'temporal.max']

P6: ['temporal.autocorr', "freq.rangesz",'region.extent', 'region.majaxis', 'region.minaxis', 'mass.region', 'threshold.area', 'freq.maxsnr.freq', 'freq.avgsnr', 'temporal.max']

P7: ['temporal.autocorr','temporal.min', 'region.extent', 'region.majmin.ratio', 'region.majaxis', 'region.minaxis', 'mass.region', 'threshold.area', 'freq.rangesz', 'temporal.max', 'spatial.min', 'mass.total']

P8: ['region.minaxis', 'threshold.area', 'mass.region','region.majmin.ratio', 'freq.rangesz', 'region.majaxis','temporal.autocorr', 'temporal.max']

P9: ['temporal.autocorr',"freq.rangesz",'region.extent', 'region.majaxis', 'region.minaxis', 'mass.region', 'threshold.area', 'freq.maxsnr.freq', 'temporal.max']

P10: ['temporal.autocorr','temporal.min','region.extent','region.majaxis','region.majmin.ratio','region.minaxis','mass.region','threshold.area','freq.rangesz', 'freq.maxsnr.freq', 'freq.avgsnr']

P11: ['temporal.autocorr', "freq.rangesz",'region.extent', 'region.majaxis', 'region.minaxis', 'mass.region', 'threshold.area', 'freq.maxsnr.freq', 'freq.avgsnr', 'temporal.max']

P12: ['spatial.min', 'region.minaxis', 'freq.rangesz','temporal.autocorr', 'threshold.area', 'region.majaxis','mass.region', 'region.majmin.ratio', 'region.extent', 'temporal.max', 'temporal.min']

P13: ['temporal.autocorr', 'region.extent', 'region.majmin.ratio', 'freq.maxsnr.freq', 'region.minaxis', 'threshold.area', 'spatial.min', 'mass.region','region.majaxis', 'freq.rangesz']

P14: ['temporal.autocorr', 'freq.rangesz','region.extent', 'region.majaxis', 'region.minaxis', 'mass.region', 'threshold.area', 'freq.maxsnr.freq', 'freq.avgsnr', 'temporal.max']

Feature list for "No Domain" data:

*These features separate signal and artifact perfectly, so they can be used individually and the accuracy will still be at 100%

• 'region.minaxis'
• 'region.majmin.ratio'
• 'region.majaxis'
• 'mass.perc'
• 'mass.region'
• 'region.orient'
• 'mass.total'
• 'region.extent'
• 'region.centroid.1'
• 'region.centroid.0'
• 'region.eccentricity'

*If they overfit for other data and do no generalize well, the feature list below can be used to get a better model. The features above were excluded from the full list while generating the list below.

['temporal.std', 'spatial.min', 'freq.rangesz', 'temporal.max', 'temporal.autocorr', 'length', 'freq.range.high', 'freq.integrate', 'freq.range.low', 'freq.avgsnr', 'freq.maxsnr.freq', 'temporal.min']

Model used: Random Forest Hyperparameters:

• n_estimators = 18
• max_features = 5

Performance with this list: Accuracy: 92.1% Precision: 89.6% Recall: 71.1%