nutmeg:ConnectivityOptions

From NITRC Wiki

• Disclaimer: explanation of Fieldtrip is just a guideline for comparison, and might be outdated. Please see http://fieldtrip.fcdonders.nl for official information

FC measure and output

• In FCM
  • Coherence (imaginary, real, absolute-value squared)
  • Phase Lag Index
  • Amplitude correlation
  • Output regardless of measure: Mean, Z-norm mean, L-image (t-test with homologous contralateral)
• In NCG
  • Coherence (imaginary, absolute-value)
  • (Straight) Amplitude correlation
    • Specify 'window' time of correlation, may be smaller than trial length, which will compute multiple sliding-window correlations
  • Chunked Amplitude correlation: like 'straight', but after HA, mean over a chunk (e.g. 100ms) of time, then correlate these chunks over time
  • Output regardless of measure: Mean
• FT for source-level
  • coherence and cross-spectral density
  • amplitude correlation and power correlation
• FT for sensor-level (freq or freqmvar)
  • same as above for source, plus the following
  • partial coherence, partialling out a channel
  • weighted phase-lag index, and debiased_PLI
  • pairwise phase consistency, and weighted_PPC
  • PLV
  • total_interdependence
  • Directed Transfer Function
  • Partial Directed Coherence
  • Phase Slope Index
  • granger (also for type 'mvar')
  • instantaneous causality (also for type 'mvar')
• FT for sensor-level (raw)
  • corr, xcorr, and spearman (although maybe code deprecated: in switch-case for these, no 'dataout' is computed)

Question specific to coherence computation: averaging

• Paper by A.P. Kulaichev, 2011 (Neurosci & Behav. Physiol.) discusses when to average over epochs/trials, amongst other things
• Kulaichev suggests to use Eq.10 rather than Eq. 9, which distinguishes between when to average in the denominator. Eq. 9 takes E[Gx(f)] * E[Gy(f)], whereas he suggests instead to use E[ Gx(f) * Gy(f)]
• Both types within Nutmeg (FCM and NCG) compute as: E[ Gxy(f) ] / E[Gx(f)] * E[Gy(f)]
• Whereas from what it seems in FT ft_connectivity_corr, that instead a third way is used: E[ Gxy(f) / Gx(f) * E[Gy(f)]
• Johanna asks: thoughts on which is best theoretically, and/or if anyone has tested this empirically? Have I misinterpreted the code?

Aspects of data to use

• FCM: explicitly choose between the following.
  • Continuous single-trial data
  • Non-event-locked multi-trial data (e.g., each trial is an artifact-free segment of a continuous recording).
  • Event-locked multi-trial data, allows running sliding sub-windows within each trial
• NCG
  • data can any number of trials, not need to make explicit whether single/multi-trial, or locked to event or not
  • can specify to use 'all' or subset of trials in dataset
  • may specify sub-windows within each trial
  • and different sub-windows allowed for different trial/condition types, as indexed by beam.s{1,1}, beam.s{1,2} etc
• FT
  • sensor level (freq, freqmvar, or mvar) (raw and timelock seem deprecated)
  • source level
  • n.b. direct call to ft_connectivityanalysis only requires cfg.method, but previous calls to set up data (e.g. ft_freqanalysis) will have specified cfg options for which trials, time-frequency windows, etc

Choice of seed voxel(s) and target voxel(s)

• FCM: choice of seed
  • All
  • selected: user can mark a region of interest (ROI) or a single voxel on a structural MRI, or enter the coordinates of a single voxel.
  • selected + contralateral
  • contralateral to selected only
  • extracerebral (e.g., EMG for cortico-muscular coherence)
• FCM: target options
  • All: Can then average over 1) All, 2) Homologous interhemisphere, 3) Heterlogous interhemispheric 4) Intrahemispheric or 5) None
  • Grid: specify spacing (every xth voxel within the original nuts.voxels coordinates)
• NCG: choice of 3 input types
  • index(indices) of voxel(s): from knowing either nuts.Lp, nuts.voxels or beam.voxels index
  • MNI coordinate(s)
  • whole brain as seed
• NCG: target
  • assumed whole brain, based on existing nuts.voxels/Lp
• NCG: using AAL atlas for 116x116 region connectivity available upon request
• FT: seed
  • cfg.refindx defines one or more indices of seed voxels
  • seems extracranial (e.g. EMG) can only be done with sensor-space metrics, not source (?) -> this is at present still done in ft_sourceanalysis/ft_sourcedescriptives, by providing cfg.refchan prior to calling ft_sourceanalysis with cfg.method = 'dics'
• FT: target
  • assumed whole brain

Statistics of output

• FCM
  • (within subjects) t-test of ispi- versus contra- to seed ROI (e.g. for tumor patients)
  • (across subjects) fcm_beam2Pimage: compares values from 1 patient against distribution of healthy-controls
  • (across subjects) fcm_corr: correlation of behavioural/clinical parameter with source-level connectivity output
• NCG
  • not available in SVN at present, but scripts exist to compute t-tests contrasting conditions over subjects per voxel, FDR corrected over voxels
• FT
  • Standard error of the mean (some metrics)