Novel Artificial Intelligence-Driven Approach for Functional Near-Infrared Spectroscopy-Based Computer-Aided Diagnosis of Major Depressive Disorder

다국어 초록 (Multilingual Abstract)

Novel Artificial Intelligence-Driven Approach for Functional Near- Infrared Spectroscopy-Based Computer-Aided Diagnosis of Major Depressive Disorder Kyeong-Gu LEE Dept. of Electronic Engineering Graduate School Hanyang University Major depressive disorder (MDD) is widely recognized as one of the most prevalent mental health disorders globally. Patients with MDD not only interfere substantially with work and relationships but also are almost 20-fold more likely to suicide than the general population. The COVID-19 pandemic and severe climate changes have gradually increased the prevalence of MDD and its emergence as a critical social issue worldwide. However, the current conventional MDD diagnosis methods based on the Diagnostic and Statistical Manual of Mental Disorders sometimes lead to misdiagnosis because these face-to-face interviews rely solely on the subjective responses of patients for diagnosis. Accordingly, the need for an accurate diagnostic system for MDD has been increasing. In response, to enhance the objectivity and reliability of the diagnosis, computer-aided diagnosis (CAD) of MDD, utilizing neuroimaging techniques, has been studied actively. Among the neuroimaging techniques, functional near-infrared spectroscopy (fNIRS) has advantages in terms of portability, cost-effectiveness, and low susceptibility to motion artifacts, compared to other modalities. In contrast, the fNIRS-based CAD studies driven by machine learning (ML) or deep learning (DL) remain relatively insufficient. To address this gap, the author performed three studies associated with ML/DL-driven methodology for the CAD of MDD. The fNIRS data of participants—48 patients with MDD and 68 healthy controls (HCs)—were obtained while they performed the verbal fluency task (VFT) and Stroop task. The preprocess of fNIRS data involved filtering, segmentation, baseline correction, and block averaging. The first study developed a DL architecture for the CAD of MDD using a convolutional neural network (CNN) with a new channel embedding layer to reflect inter- hemispheric asymmetry in prefrontal hemodynamic responses between patients with MDD and HCs, which have been known to be a distinct characteristic in previous MDD studies. The six representations for hemodynamic responses calculated from the preprocessed fNIRS data were used as inputs to the proposed CNN model—comprising three 1D depth-wise convolutional layers—with an ensemble architecture. The performance of the proposed model was evaluated using a leave-one-subject-out cross- validation strategy and compared with those of conventional machine learning and CNN models. The proposed model exhibited a high accuracy, sensitivity, and specificity of 84.48%, 83.33%, and 85.29%, respectively within the Stroop task paradigm. For the VFT paradigm, the proposed model showed an accuracy of 80.17%, with sensitivity and specificity rates at 72.92% and 85.29%, respectively. The proposed DL models can differentiate between the patients with MDD and HCs more accurately than the conventional models, demonstrating its applicability in fNIRS-based CAD systems. The second study introduced the novel XAI system tailored for the fNIRS-based CAD of MDD. The objective was to identify whether the feature extraction strategy proposed in the first study effectively captured the temporal hemodynamic activation and inter-hemispheric asymmetry characteristic of patients with MDD. Changes in the concentrations of oxygenated and deoxygenated hemoglobin, stacked in the channel direction of a 2-dimensional convolutional filter, were used as inputs of a streamlined version of the DL model proposed in the first study. Layer-wise relevance propagation (LRP) was used to obtain relevance scores across each fNIRS channel location. This study not only identified the channel locations from which dominant features were extracted but also provided a rationale for which these channels were able to identify important features. The author demonstrated that the spatial distribution of relevance scores aligned closely with those of Fisher's scores, which are indicative of the discrimination power of hand-crafted features. Moreover, through the waveform analysis for symmetric channel pairs exhibiting high relevance scores, the author identified the inter-hemispheric asymmetry, consistent with the findings of previous studies. This study demonstrated more profound insights into the decision-making mechanisms of the DL architecture, offering a more comprehensive analysis than a previous XAI study for the CAD of MDD. The third study proposed the meta-classification methodology to enhance the diagnostic accuracy of MDD. The six features were calculated from four preprocessed representations of the hemodynamic response in each fNIRS channel. Then, incorporating the extraction of features associated with the inter-hemispheric asymmetry referring to the findings from the first and second studies, the author enhanced the classification accuracies of machine learning models for the Stroop task and VFT. As a result, the classification accuracies of single classifiers for the VFT and Stroop task were 79.31% and 80.17%, respectively, demonstrating that the performances were higher than those of the first and second studies. Furthermore, meta-classifiers were implemented by combining single classifiers with the highest accuracy for each cognitive task. The meta- classification method proposed in this paper showed the highest classification accuracy of 89.66%. This study showed that the novel meta-classification method combining single classifiers for cognitive tasks could enhance the classification performance between the MDD and HC groups. The author believes that the newly proposed AI-driven approaches for the fNIRS- based CAD of MDD have enhanced both the quantity and quality of clinical investigations, thereby making a substantial contribution to advancing the fNIRS-based CAD of MDD. Nonetheless, there is a need for the collection of more extensive fNIRS data to enhance the reliability and applicability of the diagnostic system. This will enable the author to optimize the proposed DL architecture, investigate the physiological insights of MDD, such as heterogeneity, from the proposed XAI system, and develop the meta- classifier with ML or DL approaches.

목차 (Table of Contents)

• Chapter 1. Introduction 21
• 1.1. Major Depressive Disorder (MDD) 22
• 1.2. Diagnosis of MDD 23
• 1.3. Evaluating MDD Using Functional Near-Infrared Spectroscopy (fNIRS) 24
• 1.3.1 Operational Principles of fNIRS 24
• 1.3.2 Comparative Benefits of fNIRS 27
• 1.3.3 fNIRS-Derived Neuroimaging Features in MDD 29
• 1.4. Machine Learning (ML) Approaches for the Diagnosis of MDD with fNIRS 30
• 1.4.1 Overview of ML 30
• 1.4.2 ML Studies for the CAD of MDD 32
• 1.4.2.1 Conventional ML Approach for the Diagnosis of MDD 34
• 1.4.2.2 Deep Learning (DL) Approach for the Diagnosis of MDD 35
• 1.5. Aim of the Dissertation 37
• Chapter 2. Commonly Utilized fNIRS Dataset 40
• 2.1. Participants 40
• 2.2. Experimental Paradigm 42
• 2.2.1 Verbal Fluency Task 42
• 2.2.2 Stroop Task 43
• 2.3. fNIRS Data Acquisition 45
• 2.4. Preprocessing 46
• Chapter 3. Exploring a Novel DL Framework for the fNIRS-Based Computer-Aided Diagnosis (CAD) of MDD 47
• 3.1. Methods 47
• 3.1.1 Proposed DL Architecture 48
• 3.1.2 Conventional ML and DL Algorithms 54
• 3.1.2.1 Regularized Support Vector Machine (rSVM) and Shrinkage Linear Discriminator Analysis (sLDA) 54
• 3.1.2.2 ShallowConvNet and EEGNet 57
• 3.1.3 Performance Evaluation 59
• 3.2. Results 62
• 3.2.1 Performances of the Proposed Model 62
• 3.2.2 Comparisons of the Proposed Model with Existing Models 64
• 3.3. Discussion 67
• Chapter 4. Novel Explainable Artificial Intelligence (XAI) System for the fNIRS-Based CAD of MDD 70
• 4.1. Methods 70
• 4.1.1 DL Model 71
• 4.1.1.1 Performance Evaluation of DL Model 73
• 4.1.1.2 Layer-Wise Relevance Propagation (LRP) 74
• 4.1.2 rSVM 76
• 4.1.2.1 Performance Evaluation of rSVM 76
• 4.1.2.2 Fisher Score Distribution 77
• 4.2. Results 78
• 4.2.1 Performance 78
• 4.2.2 Relevance Score Distribution 79
• 4.2.3 Distribution of Fisher Score 85
• 4.3. Discussion 87
• Chapter 5. Meta-classification for the CAD of MDD Using fNIRS Data Collected During Multiple Cognitive Tasks 96
• 5.1. Methods 96
• 5.1.1 Feature Extraction 97
• 5.1.2 Classification 97
• 5.1.2.1 Validation Procedure 97
• 5.1.2.2 Single Classifier 99
• 5.1.2.3 Meta-Classification 99
• 5.2. Results 101
• 5.2.1 Behavior Analysis 101
• 5.2.2 Classification Performance 102
• 5.3. Discussion 103
• Chapter 6. Conclusion 108
• Reference 112
• 국문 요지 124
• 감사의 글 127