Development of Electroencephalogram (EEG) Based Neurophysiological Biomarkers and Computer-Aided Diagnostic System for Psychiatric Disorders

Abstract

Post-traumatic stress disorder (PTSD) develops after traumatic event, they demonstrated by unique symptoms such as hyperarousal, re-experiencing, and avoidance. Major depressive disorder (MDD) represents several symptoms involving loss of interest and depressive mood, and they also shows declined cognitive functions. PTSD and MDD patients have similar symptoms such as dysphoria and numbing, this problem causes confusions when diagnoses in each mental disorder. To solve this problem, the necessity of objective bio-markers to assist precise diagnosis in each disorder is continuously proposed. The purpose of this dissertation is development of 2-stage computer-aided diagnosis (CAD) system using neurophysiological biomarkers for diagnosis of PTSD and MDD patients. First study is the development of biomarker for understanding of pathophysiology of PTSD. In this study, to investigate the cortical network of patients with PTSD, source-level weighted brain networks were investigated at both global and nodal levels. Sixty six cortical source signals were estimated from 78 PTSD and 58 healthy controls (HCs) using resting-state EEG. Four global indices (strength, clustering coefficient, path length, and efficiency), and nodal clustering coefficient were evaluated in six frequency bands (delta, theta, alpha, low beta, high beta, and gamma). PTSD showed decreased global strength, clustering coefficient, and efficiency, in delta, theta, and low beta band, and enhanced PL in theta and low beta band. In low beta band, the strength and clustering coefficient correlated positively with the anxiety scores, while path length had a negative correlation. Additionally, nodal clustering coefficients were reduced in PTSD in delta, theta, and low beta band. Nodal clustering coefficients of theta band correlated negatively with rumination and re-experience symptom scores; while, nodal clustering coefficients in low beta band correlated positively with anxiety and pain severity. Inefficiently altered and symptom-dependent changes in cortical networks were seen in PTSD. Our source-level cortical network indices might be promising biomarkers for evaluating PTSD. Secondly, as with PTSD patients, to investigate the cortical network of patients with MDD, source-level weighted brain networks were investigated at both global and nodal level. 87 MDD and 58 healthy controls were recruited. Four global indices (strength, clustering coefficient, path length, and efficiency), and two nodal index (nodal clustering coefficient and eigenvector centrality) were evaluated in six frequency bands. MDD showed decreased global strength and clustering coefficient in theta and alpha band, and decreased efficiency and enhanced path length were found in alpha band. Additionally, altered nodal indices were found only in alpha beta band. Enhanced eigenvector centrality positively correlated with depressive and anxiety scores; reduced nodal clustering coefficient negatively correlated with depressive and anxiety scales. Altered network indices reflected abnormal emotional information processing of MDD during resting-state. Thirdly, to investigate cognitive function of PTSD and MDD, P300 event-related potential (ERP) study was performed in both sensor-level and source-level. EEG signals were recorded from 53 PTSD, 77 MDD and 39 HCs during performing auditory oddball task. P300 amplitude and latency were estimate for sensor-level, and current source densities were estimated by sLORETA. As a result, PTSD showed significantly reduced P300 amplitude as compared to HCs and MDD, and latency of PTSD was significantly prolonged as compared to MDD. Moreover, PTSD showed significantly reduced current source densities, particularly, sources in lingual gyrus, anterior cingulate cortex, parahippocampal gyrus, and inferior frontal gyrus significantly correlated with symptom scores. Altered P300 characteristics of PTSD reflected declined cognitive function such as automatic processing compared to MDD, and relationships between abnormal sources and symptom scores might support cognitive deficits of PTSD. Finally, the author differentiated PTSD from MDD based on machine learning using developed P300 biomarkers. Three classification models (PTSD-HCs, MDD-HCS, and PTSD-MDD) were built. P300 amplitudes and latency of 62 electrodes (total 124) were evaluated for sensor-level features, and 15 source-level features estimated from source activity using sLORETA. The 1 to 15 candidate features were selected using Fisher’s score from sensor, source and combined feature sets, and the classification accuracy was evaluated. The maximum classification accuracy and selected feature set when achieved maximum accuracy in each classification model were as follows: 1) PTSD-HCs: accuracy – 79.35 % selected features – source-level feature set (n = 6); 2) MDD-HCs: accuracy – 71.55 %, selected features – combined feature set (sensor = 12; source = 2); 3) PTSD-MDD: accuracy – 74.92 %, selected features – combined feature set (sensor = 2; source = 4). In this dissertation, the author developed 2-stage CAD system using EEG based neurophysiological biomarkers for diagnosis of PTSD and MDD patients. The EEG based neurophysiological biomarkers may help to understand pathophysiology of PTSD and MDD and be used as treatment predictors. Also, the proposed 2-stage CAD system may be a promising diagnostic system for diagnosis of PTSD and MDD.