Biomarker Discovery for Mental Disorders from Neuroimaging Data

Abstract

A mental disorder is a medical condition that disrupts a person's thinking, feeling, mood, ability to relate to others and daily functioning. Despite decades of research, the exact cause of most mental disorders remains unknown and no objective tests are available for their diagnosis. It is crucial to explore quantitative, discriminative and interpretable biomarkers for mental disorders, which could not only assist in clinical diagnosis, but also help in gaining insight into the underlying mechanisms of the illnesses. In this research, we explore neuroimaging biomarkers for two common mental disorders: schizophrenia and borderline personality disorder (BPD). An interdisciplinary research framework covering feature extraction, selection, classification and validation is presented. Signal processing and graph theory approaches are employed to process and model neuroimaging data, and create meaningful feature sets to characterize brain activity, connectivity and network topologies. Machine learning based feature ranking and classification is performed to select discriminating feature subset and distinguish psychiatric patients from healthy subjects. Statistical analysis is performed to validate the significance of the identified features and control the False Discovery Rate (FDR). In the first part of the dissertation, we explore spatial-temporal-spectral neural oscillation patterns for schizophrenia using magnetoencephalography (MEG) data. We first extract Event-Related Desynchronization/Synchronization (ERDS) patterns along the space, time and frequency dimensions combined. A two-step feature ranking algorithm combining F-score filtering and Support Vector Machine – Recursive Feature Elimination (SVM-RFE) algorithm is applied to select a small subset of features according to their discriminating power. With top 20 ERDS features, 90% specificity and 91.67% sensitivity is achieved in classifying 12 schizophrenia patients from 10 healthy controls using a linear SVM classifier, following cross validation procedure. Next, two novel spatial-temporal-spectral feature sets, the Band Power Ratio (BPR) and the Window Power Ratio (WPR) are created, based on the Power Spectral Density of MEG data. Cluster-based nonparametric permutation tests are employed to identify key features with significant between-group difference, which control the FDR while maintaining low False Negative Rate. The minimum-Redundancy-Maximum-Relevance criteria are then employed to select the optimal feature combinations for classification. Based on only 2 WPR and 1 BPR feature combined, over 95% cross validation classification accuracy is achieved using three different linear classifiers separately, which indicates strong discriminating power of the key Spectral Power Ratio features. Using spectral power features, a computer-aided schizophrenia screening system based on majority voting of single MEG trials is then presented. In the second part of the dissertation, we explore functional brain network connectivity and topology patterns for BPD using resting-state functional magnetic resonance imaging (fMRI) data. Frequency-specific brain networks are constructed by correlating wavelet-filtered fMRI time series from 82 cortical and subcortical regions. Network-based statistics are employed to identify altered connections using cluster-based thresholding of statistical maps. An interconnected subnetwork in 0.03–0.06Hz frequency band is identified that shows significantly lower connectivity strength in patients. The mean connectivity of the subnetwork shows negative correlation with several key clinical symptom scores, and achieves 90% sensitive and 90% specificity in BPD classification using a simple linear classifier. We further employ graph theory to investigate the global and local topological structure of the frequency-specific brain connectivity networks. Statistical analysis show that BPD patients have significantly larger measures of global network topology, including the size of largest connected component, clustering coefficient, small-worldness and local efficiency, indicating increased local cliquishness in the functional brain network. These global topology metrics show positive correlations with several clinical symptom scores associated with BPD. Additionally, compared to controls, patients show lower nodal centrality at several hub regions but higher centrality at several non-hub regions in the network. These findings may add to the current understanding of functional brain networks in BPD.