FBIRN:NeuroInfStatsTuesAM2006

Data Mining

• History of Data Mining
  • 1960's Tukey
  • Predictive Modeling (CART)
  • Machine Learning (CS)
  • Relational Databases
  • Web
  • Hardware
• Convergence of above technologies has lead to data mining
  • There is a lot of hype regarding data mining
• FBIRN has a small N in the data mining sense
  • Clustering, look for subgroups in the schizophrenia patients
  • fBIRN data has a rich structure
  • Focused on data analysis right now, do algorithm comparisons
  • Make sure that the raw spatial data, timecourse data is available
• Standardized data for data analyses
  • MRI Machine --> "raw data" --> intermediate data --> Beta-maps, etc
    • inter-subject issues in preprocessing
• data analysis recommendations
  • give someting standard and consistent across sites
  • should data vetting be performed, runs or subjects tossed based on results of first level analysis
  • time series of variance, auto-correlation
  • having standard datasets available so other disciplines (e.g., computer scientists, statisticians) can look at the data
  • proposal, FIPS 1.0 produces standard results (beta-maps) on phaseII, then spend 6 months improving the pipeline
  • make data available after each major preprocessing step
    • after pre-whitening/smoothing
    • after motion correction (would need to modify FIPS/FEAT)
    • after slice timing correction (would need to modify FIPS/FEAT)
    • currently FIPS final products create the output of FEAT
  • phaseII
    • uncompressed raw data is just over 1Gb/subject visit
    • intermediate time series 20Gb/subject visit
    • may be problematic to store all of the intermediate steps, due to disk space and downloading times
• what can we do for every subject?
• what can we do on a subset of subjects?
• FIAC
  • standard data products
  • 1st level
    • raw
    • motion corrected
      • motion parameters
    • slice-time corrected
    • meta-data, details of the paradigm
      • best practices, what we think the design matrix should be
      • what happened at what time, how long the blocks were, etc.
  • 2nd level
    • constrast copes in standard space
    • varcopes in standard space
    • meta-data,
      • what effect
      • degrees of freedom
      • smoothness
      • T image
      • threshold
• Standard Data Products Plan
  • Establish a working group
  • Establish the requirements of what is released besides raw data and when
  • QA reports on raw and on intermediate data products
• Data QA
  • Variance of image over time series
  • Variacne oc (Xt-Xt+1) over time series
    • scale median to 100
  • Good for residuals
    • (Outlier count per image / Expected outlier count ) * 100 for each image yields a time series
    • Normality test at each voxel, look at how many p values smaller than 0.05
      • (Number of significant voxels / Expected significant voxels) * 100