[Paper Review] Classification of EEG-Based Brain Connectivity Networks in Schizophrenia Using a Multi-Domain Connectome Convolutional Neural Network

Chun‐Ren Phang, Chee‐Ming Ting+2 more|arXiv (Cornell University)|Mar 21, 2019

Functional Brain Connectivity Studies48 citations

TL;DR

This paper introduces MDC-CNN, a multi-domain connectome CNN that fuses time-domain VAR, frequency-domain PDC, and topological CN features from EEG to classify schizophrenia vs. healthy controls, outperforming SVM and single-domain CNNs.

ABSTRACT

We exploit altered patterns in brain functional connectivity as features for automatic discriminative analysis of neuropsychiatric patients. Deep learning methods have been introduced to functional network classification only very recently for fMRI, and the proposed architectures essentially focused on a single type of connectivity measure. We propose a deep convolutional neural network (CNN) framework for classification of electroencephalogram (EEG)-derived brain connectome in schizophrenia (SZ). To capture complementary aspects of disrupted connectivity in SZ, we explore combination of various connectivity features consisting of time and frequency-domain metrics of effective connectivity based on vector autoregressive model and partial directed coherence, and complex network measures of network topology. We design a novel multi-domain connectome CNN (MDC-CNN) based on a parallel ensemble of 1D and 2D CNNs to integrate the features from various domains and dimensions using different fusion strategies. Hierarchical latent representations learned by the multiple convolutional layers from EEG connectivity reveal apparent group differences between SZ and healthy controls (HC). Results on a large resting-state EEG dataset show that the proposed CNNs significantly outperform traditional support vector machine classifiers. The MDC-CNN with combined connectivity features further improves performance over single-domain CNNs using individual features, achieving remarkable accuracy of $93.06\%$ with a decision-level fusion. The proposed MDC-CNN by integrating information from diverse brain connectivity descriptors is able to accurately discriminate SZ from HC. The new framework is potentially useful for developing diagnostic tools for SZ and other disorders.

Motivation & Objective

• Motivate automatic discrimination of schizophrenia using altered EEG-based brain connectivity patterns.
• Investigate whether combining multiple connectivity domains improves classification over single-domain approaches.
• Develop a deep learning architecture that can fuse time-domain, frequency-domain, and topological connectivity features.
• Evaluate MDC-CNN against traditional classifiers and single-domain models on resting-state EEG data.

Proposed method

• Extract directed connectivity features from multi-channel EEG using VAR coefficients (time-domain) and PDC (frequency-domain).
• Compute complex network measures (degree, global efficiency, clustering, transitivity) from connectivity matrices as topology features.
• Develop MDC-CNN with two 2D-CNNs for VAR and PDC matrices and one 1D-CNN for CN features in parallel.
• Explore three fusion strategies: feature-level fusion, score-level fusion, and decision-level fusion.
• Train the model with Adam optimizer, 500 epochs, dropout in fully-connected layers, and cross-entropy loss for SZ vs. HC classification.

Experimental results

Research questions

• RQ1 Can EEG-based directed connectivity measures discriminate schizophrenia from healthy controls?
• RQ2 Does integrating time-domain, frequency-domain, and topological connectivity features improve classification accuracy?
• RQ3 Which fusion strategy yields best performance for multi-domain connectome classification?
• RQ4 How does MDC-CNN compare to SVM and single-domain CNN approaches on resting-state EEG data?

Key findings

• MDC-CNN with combined features achieves the best accuracy of 92.87% (modified 93.06%) using decision-level fusion.
• Time-domain VAR features alone yielded high accuracy (CNN: 90.59%; all bands with SVM: 88.19% modified), outperforming frequency-domain PDC in several settings.
• CN topological features alone performed worse than VAR or PDC in most bands but provided complementary information.
• Fusion of VAR, PDC, and CN via majority voting outperformed single-domain CNNs and SVM across all feature sets.
• All-band PDC-based CNNs achieved up to 89.19% accuracy with 93.33% sensitivity and 93.33% specificity in some configurations.
• Decision-level fusion consistently offered the best trade-off between accuracy and robustness among fusion strategies.