The aim of this study was to develop and validate a method of disease classification for bipolar disorder (BD) by functional activity and connectivity using radiomics analysis. Ninety patients with unmedicated BD II as well as 117 healthy controls underwent resting-state functional magnetic resonance imaging (rs-fMRI). A total of 4 types of 7018 features were extracted after preprocessing, including mean regional homogeneity (mReHo), mean amplitude of low-frequency fluctuation (mALFF), resting-state functional connectivity (RSFC), and voxel-mirrored homotopic connectivity (VMHC). Then, predictive features were selected by Mann-Whitney U test and removing variables with a high correlation. Least absolute shrinkage and selection operator (LASSO) method was further used to select features. At last, support vector machine (SVM) model was used to estimate the state of each subject based on the selected features after LASSO. Sixty-five features including 54 RSFCs, 7 mALFFs, 1 mReHo, and 3 VMHCs were selected. The accuracy and area under curve (AUC) of the SVM model built based on the 65 features is 87.3% and 0.919 in the training dataset, respectively, and the accuracy and AUC of this model validated in the validation dataset is 80.5% and 0.838, respectively. These findings demonstrate a valid radiomics approach by rs-fMRI can identify BD individuals from healthy controls with a high classification accuracy, providing the potential adjunctive approach to clinical diagnostic systems.
Keywords: bipolar disorder; machine learning; radiomics; resting-state functional magnetic resonance imaging.
© The Author(s) 2019. Published by Oxford University Press. All rights reserved. For permissions, please e-mail: [email protected].