In recent years, more and more attention has been paid to the individual level classification and prediction of psychiatric disorders based on neuroimaging. Using a variety of neuroimaging modalities such as functional magnetic resonance (fMRI), structural magnetic resonance (sMRI) and diffusion magnetic resonance imaging (dMRI), along with machine learning techniques, more than hundreds of studies have been carried out on the accurate classification of brain disorders. However these classification studies still have some disadvantages, such as biased feature selection, over fitting, small sample size, and poor generalization. Therefore, based on the summary of previous studies, this paper improved the support vector machine (SVM) classification algorithm, which was the most widely used method in psychiatric disorder classification. Combined with independent component analysis (ICA), subspace similarity, feature selection, multimodal fusion, multi-kernel classification and other machine learning technologies, the proposed methods were performed on the classification of different psychiatric disorders, and the classification results and the selected discriminative brain regions were further analyzed. The main achievements of our study are as follows:
(1) A classification algorithm of support vector machine based on subspace of independent components was proposed. Firstly, the individual level spatial independent component (IC) of each subject was extracted, and then the similarity matrix between the subspaces spanned by different subjects’ ICs was constructed by using the specific subspace distance. The similarity matrix was further partitioned in the kernel space. Finally, the similarity matrix was embedded into SVM classifier as kernel function to classify subjects and simultaneously to select the corresponding discriminative independent component combination. This method was applied to the early differential diagnosis of patients with bipolar disorder (BD) and major depression disorder (MDD). The accuracy of classification was as high as 92.4%, and five maximally contributory IC brain networks were identified. These intrinsic connection networks (ICNs) included default mode network, salience network, dorsal attention network, frontoparietal central executive network, subcortical regions including caudate body, thalamus and parahippocampal gyrus. Based on networks mentioned above, 12 complicated patients with unclear mood disorders were classified as possible groups (BD/MDD) and compared with the corresponding medication-class of response, and the blind test accuracy was 91.7%, demonstrating this method had broad clinical application prospects.
(2) A multi-kernel SVM classification algorithm using uniform group level reference signal was proposed. Based on the algorithm proposed above, the uniform group level reference signal template calculated by large samples of healthy controls was adopted to extract individual level ICs. For each independent component of specific brain network, subspace similarity method was adopted to construct the kernel of the brain network, and then multi-kernel SVM was adopted to combine different brain networks for classification. In the classification of bipolar disorder (BD) and schizophrenic (SZ) patients from multiple sites, the superiority of multi-kernel classification was verified, and the discriminative brain regions between BD and SZ patients were found located in cognitive control network, subcortical regions, visual network and cerebellum. This method provided a unified template for the comparison of independent components among different sites. Comprehensive experiments of multi-kernel / subspace, all / functional domains, number of components and linear / nonlinear comparison were carried out. Under the optimal experimental condition, the classification accuracy was 80.2%, and the overlapping discriminative components among different sites were found.
(3) A multi-kernel SVM classification algorithm for multi-modal features was proposed. Firstly, the group level corresponding independent components of different modalities were extracted by multi-modal fusion method (mCCA + jICA) to make masks for different modal features. Then, the subjects were divided by nested 10 fold cross validation to select features and components. The subspace similarity between different subjects based on selected feature components was used as the kernel of each modality. Finally, multi-kernel SVM was used to combine each modality for classification. Two independent multi-modal datasets were used to classify the SZ patients and healthy controls (HC). Experiments demonstrated that the accuracy of multi-modal methods were 6-30% higher than that of single-modal method. The discriminative independent component brain networks selected from the two datasets were also correspondent to a certain extent.
(4) The performance of multi-site classification of SVM based on subspace of ICs was verified. Firstly, in each site, the method proposed in this paper was used to extract group level ICs, select discriminative components and build classification models. Then in other sites, the group level ICs were used as reference signals to extract individual level ICs and the models built before were adopted for classification test. This method was applied to the multi-site classification of ADHD patients and HCs from two sites. ADHD is a kind of psychiatric disorders which is difficult to predict and classify, and the classification accuracies are mostly 60-70% in existing studies. The cross-site classification of the proposed method was verified. (The accuracies of single site classification were both more than 80%, and those of multiple sites prediction were both more than 70%.) The abnormal changes of functional brain networks in ADHD patients including attention network, default mode network, control network, caudate and thalamus were found, indicating the promising generalization.
In this paper, a series of classification algorithms based on MRI, especially fMRI, of SVM based on subspace of ICs were proposed, and combined with different machine learning technologies including group information guided independent component analysis, multi-kernel technique, multi-modal fusion etc. These methods were applied to different psychiatric disorder classification, to verify the effectiveness of the proposed methods in independent blind testing, multi-modal and multi-site classification and to provide the basis of exploring the potential abnormal biomarkers of various psychiatric disorders.