Trends in multivariate group analysis of neuroimaging data

Based on a talk held at Neurospin, Saclay, France, November, 2014

Transcript

1. Trends in multivariate group analysis and cross- subject decoding Denis A. Engemann
2. our goal today contrast strategies that address the problem of between-subject variability in group analysis
3. between-subject variability: an ancient problem
4. [Procustes] was a rogue smith and bandit from Attica who physically attacked people by stretching them or cutting off their legs, so as to force them to fit the size of an iron bed. In general, when something is Procrustean, different lengths or sizes or properties are fitted to an arbitrary standard.
5. Prcrustes Analysis Match objects as closely as possible by applying optimal • translation • scaling • rotation
9. Hyper alignment • Key tool: projection to common space / model • Availability: Matlab and Python • Modality: For now only used in concert with fMRI
11. Competition: decoding across subjects Olivetti E, Kia SM, Avesani P (2014), "MEG decoding across subjects", Pattern Recognition in Neuroimaging, 2014 International Workshop on, pp.1-4, 4-6 June 2014. doi: 10.1109/PRNI.2014.6858538
12. How competitive exactly?
13. temporal between-subject variability
14. subject 01 subject 02 subject 03 subject 04 subject 05 subject 06 spatial between-subject variability
15. Multiclass Brain-Computer Interface Classification by Riemannian Geometry Alexandre Barachant, St ́ephane Bonnet, Marco Congedo, Christian Jutten 1. Spatial filters enhancing the SNR class-wise (8 virtual signals) 2. New trials based on incoming and averaged evoked signals 3. Spatial covariance between ensuing signals 4. Apply Riemannian geometry as descriptor: (R distance, mean, tangent mapping) 5. Map to eucledian space and fit + classify
16. Spatial filter +Riemann Geometry • Key tool: compression of individual spatial and temporal characteristics • Availability: Matlab (so far …). • Modality: For now only used in concert with EEG, MEG in offline and BCI contexts
17. Automated model selection in covariance estimation and spatial whitening of MEG and EEG signals. Engemann, D. A. and Gramfort, A. (in press). Neuroimage
18. M inimum N orm E stimates aka Tikhonov Regularization aka Ridge Regression •constraint linear model (cf. beamformer, S-LORETA, ...) •Gaussian, uncorrelated noise •whitening via covariance ˆ X = RGt(GRGt + C) 1Y ˆ X = R ˜ Gt( ˜ GR ˜ Gt + I) 1 ˜ Y unwhitened whitened 98.749% M/EEG users used whitening ˆ X = RGt(GRGt + C) 1Y ˆ X = R ˜ Gt( ˜ GR ˜ Gt + I) 1 ˜ Y
19. 1. Hand-set (REG) 2. Ledoit-Wolf (LW) 3. Cross-validated shrinkage (SC) 4. Probabilistic PCA (PPCA) 5. Factor Analysis (FA) Compare estimators using CV simple, fast complex, slow CPPCA = HHt + 2IN CLW = (1 ↵)C + ↵µI CSC = (1 ↵)C + ↵µI CSC = (1 ↵)C + ↵µI CLW = (1 ↵)C + ↵µI C0 = C + ↵I, ↵ > 0 CP P CA = HHt + 2IN CF A = HHt + diag( 1 , . . . , D)
20. faces > scrambled SPM faces dataset, Henson (2003) 20 epochs 40 epochs 60 epochs worst estimator best estimator std
21. faces > scrambled SPM faces dataset, Henson (2003) 20 epochs 40 epochs 60 epochs std The regularized covariance stabilizes dSPM source amplitudes worst estimator best estimator
22. Automated regularization • Key tool: learn regularization parameters from the data • Availability: Python. • Modality: For now only used in concert with EEG, MEG in offline contexts
23. What have we learned? • Strategy 1: transform data to a common feature space • Strategy 2: Condense, compress individual spatial and temporal features in dedicated e.g. geometrical metrics and just learn on a few features. • Stragey 3: Reduce sample bias and estimation variance by learning regularizaiton parameters from the data