Identification of bipolar disorder using a combination of multimodality magnetic resonance imaging and machine learning techniques

Abstract Background Bipolar disorder (BPD) is a common mood disorder that is often goes misdiagnosed or undiagnosed. Recently, machine learning techniques have been combined with neuroimaging methods to aid in the diagnosis of BPD. However, most studies have focused on the construction of classifiers based on single-modality MRI. Hence, in this study, we aimed to construct a support vector machine (SVM) model using a combination of structural and functional MRI, which could be used to accurately identify patients with BPD. Methods In total, 44 patients with BPD and 36 healthy controls were enrolled in the study. Clinical evaluation and MRI scans were performed for each subject. Next, image pre-processing, VBM and ReHo analyses were performed. The ReHo values of each subject in the clusters showing significant differences were extracted. Further, LASSO approach was recruited to screen features. Based on selected features, the SVM model was established, and discriminant analysis was performed. Results After using the two-sample t-test with multiple comparisons, a total of 8 clusters were extracted from the data (VBM = 6; ReHo = 2). Next, we used both VBM and ReHo data to construct the new SVM classifier, which could effectively identify patients with BPD at an accuracy of 87.5% (95%CI: 72.5–95.3%), sensitivity of 86.4% (95%CI: 64.0–96.4%), and specificity of 88.9% (95%CI: 63.9–98.0%) in the test data ( p = 0.0022). Conclusions A combination of structural and functional MRI can be of added value in the construction of SVM classifiers to aid in the accurate identification of BPD in the clinic.