Combining Multiple Resting-State fMRI Features during Classification: Optimized Frameworks and Their Application to Nicotine Addiction

Xiaoyu Ding,Yihong Yang,Elliot A Stein,Thomas J Ross

doi:10.3389/fnhum.2017.00362

Xiaoyu Ding, Yihong Yang + Show 2 more

Open Access

https://doi.org/10.3389/fnhum.2017.00362

Copy DOI

Abstract

Machine learning techniques have been applied to resting-state fMRI data to predict neurological or neuropsychiatric disease states. Existing studies have used either a single type of resting-state feature or a few feature types (<4) in the prediction model. However, resting-state data can be processed in many different ways, yielding different feature types containing complementary and/or novel information, leaving uncertain the most informative features to provide to the classifier. In this study, multiple resting-state features were calculated from two main analytical categories: local measures and network measures. Feature selection was adopted using an optimized grid-search approach selecting top ranked features from statistical tests. We then tested three optimized frameworks: feature combination, kernel combination, and classifier combination, all using the support vector machine as an elementary classifier, to combine these resting-state feature types. When applied to nicotine addiction, with a cohort size of 100 smokers and 100 non-smokers, via a 10-fold cross-validation procedure, the feature combination and the classifier combination achieved an accuracy of 75.5%, while the kernel combination achieved a 73.0% accuracy; all three combination frameworks improved classification performance compared to the single feature type based results (best accuracy 70.5%). This study not only reveals the discriminative power of resting-state data, but also demonstrates the efficiency of combining multiple features from one data phenotype to improve classification performance.

Highlights

Machine learning techniques are playing an increasingly important role in neuroscience research to explore various brain functions (Klöppel et al, 2012; Richiardi et al, 2013; Sundermann et al, 2014; Gabrieli et al, 2015)
Using nicotine dependence as a model system and using all feature types, the three approaches overall yielded very similar results; the classifier combination and the feature combination frameworks reached an accuracy of 75.5%, while the kernel combination achieved a 73.0% accuracy
As a single feature type, Temporal Correlation (TC) achieved the highest accuracy of 70.5%, but most of the other feature types individually only performed at slightly above chance with the exception of Granger Causality (GC)

Summary

Introduction

Machine learning techniques are playing an increasingly important role in neuroscience research to explore various brain functions (Klöppel et al, 2012; Richiardi et al, 2013; Sundermann et al, 2014; Gabrieli et al, 2015) They have been applied to neuroimaging data to predict group membership, which may lead to brain-based biomarkers of disease (Chen and Herskovits, 2010; Wang et al, 2010; Zhang and Shen, 2012; Hart et al, 2014; Pariyadath et al, 2014; Ding et al, 2015; Jie et al, 2015; Libero et al, 2015; Liu et al, 2015; Moradi et al, 2015; Suk et al, 2015; Arbabshirani et al, 2016). A commonly adopted kernel function is the Gaussian radial basis function (RBF) that maps the input samples into a Hilbert space, corresponding to a non-linear SVM called RBF kernel SVM (RBF-SVM) (Burges, 1998)

Methods

Results

Discussion

Conclusion