Inter-subject pattern analysis
A powerful scheme for group-level MVPA
Sylvain Takerkart
with Qi Wang, Bastien Cagna, Thierry Chaminade
http://int.univ-amu.fr
http://meca-brain.org
http://neuralbasesofcommunication.eu
Outline
- Group-level MVPA
- Benchmark on real data
- Insights from artificial data
- Conclusions
Multivariate pattern analysis
predictive model: f(X) = y
assessing neural coding principles
[Varoquaux 2014]
[Haynes 2015]
X: the brain pattern, y: the target variable
Multivariate pattern analysis
"traditionally" performed within-subject
"Here, we show that the topographic arrangement of the full pattern
of response is consistent within subjects,
...but we are not able to
perform a similar correlational analysis across subjects"
[Haxby 2001]
Inter-individual variability
Methods to overcome variability
[Haxby 2011] [Grosenick 2013] [Takerkart 2014] [Yamada 2016] [Hoyos-Idrobo 2018] etc.
...at the group level
This tests...
...is consistent across subjects
...how much decoding...
...how coding...
Benchmarking G-MVPA and ISPA
Everything is identical, except the cross-validation
Benchmarking G-MVPA and ISPA
- binary classification task
- logistic regression
- permutation test on accuracies at the second level
- same number of observations because of LOSO
Everything identical, except the cross-validation
Outline
- Group-level MVPA
- Benchmark on real data
- Insights from artificial data
- Conclusions
Searchlight decoding
[Kahnt 2011]
Mapping local multivariate effects using sliding window decoding
Dataset 1
- decoding right hand vs left hand tapping
- 15 subjects
- 360 trials / subject
Dataset 2
- decoding vocal vs. non-vocal audio stimuli
- 39 subjects
- 144 trials / subject
Dataset 1
Dataset 2
Dataset 1
Dataset 2
Dataset 1
Dataset 2
Conclusion (real data)
- results are consistent, but with differences!
- ISPA "detects more" on Dataset 1
- G-MVPA "detects more" on Dataset 2
Why?
Outline
- Group-level MVPA
- Benchmark on real data
- Insights from artificial data
- Conclusions
Designing artificial data
- 21 subjects
- 2 conditions
- 100 points (trials) per condition per subject
- total = 4200 points per dataset
- 2D patterns
- parametrically controlled characteristics
Effect size
Inter-individual variability
Some example datasets...
Experiments
- 2 parameters : effect size, variability
- 100 data sets for each value of the 2 parameters
- evaluation criterion: the number of data sets which offer significant group decoding ?
G-MVPA results
Number of datasets with significant decoding
ISPA results
Number of datasets with significant decoding
Comparing G-MVPA and ISPA
Outline
- Group-level MVPA
- Benchmark on real data
- Insights from artificial data
- Conclusions
Take-home message #1
Machine learning for neuroimaging...
...small sample size regime
Training set size matters!
Designing experiments with more trials...
Take-home message #2
Be careful when interpreting G-WSPA results!
Sources of these differences?
[Todd 2013] [Allefeld 2016] [Gilron 2017]
- bad registration?
- functional variability?
- idiosyncrasies?
Take-home message #3
Interpretation is unambiguous with ISPA
...at no extra computational cost...
ISPA can be more powerful
Q. Wang, B. Cagna, T. Chaminade, and S. Takerkart, ‘Inter-subject pattern analysis: A straightforward and powerful scheme for group-level MVPA’, NeuroImage, vol. 204, Jan. 2020.
paper:
https://hal.archives-ouvertes.fr/hal-02313840/
nilearn-based code:
https://github.com/SylvainTakerkart/inter_subject_pattern_analysis
data:
https://openneuro.org/datasets/ds001771/
https://zenodo.org/record/2591038
slides:
https://slides.com/takerkart/
Go see Qi Wang's poster!
Extras...
Dataset 1
Dataset 2