Objective: Investigate MVPA and EL for understanding the relationship between structural and functional abnormalities in AOIFD.

Background: AOIFD is a movement disorder with diverse phenotypes and unknown pathophysiology. The majority of neuroimaging studies have relied on univariate analyses of unimodal data for detection of regional abnormalities. However, as the relationship between structural and functional abnormalities in AOIFD is usually complex, these traditional analytic models become ineffective for abnormality integration. MVPA and EL may help exploit the complementarity of information from multimodal data.

Method: Structural (s) and resting-state (rs) functional MRI data were analysed from age-and sex-matched cohorts of 70 AOIFD patients (10 cervical, 10 blephrospasm, 15 laryngeal, 17 musician’s and 17 writer’s cramp) and 70 healthy controls. Regional homogeneity and gray matter volumetric maps were extracted from the preprocessed data. Two masks were employed: 1) an a-priori mask consisting of regions found abnormal in AOIFD literature, and 2) a whole-brain mask. Following feature selection via principal component analysis, random forest and support vector machine classifiers were then applied to the feature vectors. Three ways of combining the multimodal data were compared: 1) early feature integration, 2) late feature integration, and 3) decision integration via EL. Ten-fold cross validation was employed to evaluate the performance of the MVPA model.

Results: The sMRI data achieved a higher classification accuracy (71.42% with the a-priori mask, 85.71% with the whole-brain mask) than the rsfMRI data (57.14% with the a-priori mask, 78.57% with the whole-brain mask). However, the combination of sMRI and rsMRI, via the decision integration approach, outperformed single MRI modality classification by reaching 85.71% accuracy (a-priori mask) and 92.85% (whole-brain mask).

Conclusion: Decision integration proved to be the most effective way of combining information from different modalities in AOIFD. When the analysis was limited to the a-priori mask, it led to lower classification performance than the whole-brain mask. This suggests that data from regions across the brain were essential when classifying patients and controls, thus supporting the multi-network concept of the disorder.

References: [1] H. A. Jinnah et al., “The focal dystonias: current views and challenges for future research,” Mov Disord, vol. 28, no. 7, pp. 926-43, Jun 15 2013. [2] R. Gilron, J. Rosenblatt, O. Koyejo, R. A. Poldrack, and R. Mukamel, “What’s in a pattern? Examining the type of signal multivariate analysis uncovers at the group level,” Neuroimage, vol. 146, pp. 113 120, Feb 1 2017. [3] T. G. Dietteric, “Ensemble Methods in Machine Learning,” International Workshop on Multiple Classifier Systems, vol. LNCS 1857, pp. 1-15, 2000. [4] T. M. Schouten et al., “Combining anatomical, diffusion, and resting state functional magnetic resonance imaging for individual classification of mild and moderate Alzheimer’s disease,” Neuroimage Clin, vol. 11, pp. 46-51, 2016. [5] L. E. Libero, T. P. DeRamus, A. C. Lahti, G. Deshpande, and R. K. Kana, “Multimodal neuroimaging based classification of autism spectrum disorder using anatomical, neurochemical, and white matter correlates,” Cortex, vol. 66, pp. 46-59, May 2015. [6] G. Battistella, P. Termsarasab, R. A. Ramdhani, S. Fuertinger, and K. Simonyan, “Isolated Focal Dystonia as a Disorder of Large-Scale Functional Networks,” Cereb Cortex, Dec 17 2015.

« Back to 2019 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/evaluating-multimodal-integration-of-abnormalities-in-adult-onset-idiopathic-focal-dystonia-aoifd-via-multivariate-pattern-analysis-mvpa-and-ensemble-learning-el/