Source-linked AI summary
Classification of EEG-Based Brain Connectivity Networks in Schizophrenia Using a Multi-Domain Connectome Convolutional Neural Network
Chun-Ren Phang, Chee-Ming Ting, Fuad Noman, Hernando Ombao
TL;DR
The paper targets the limited use of network-level and multi-domain connectivity information in EEG-based schizophrenia classification. It introduces a multi-domain connectome CNN that fuses directed connectivity and topology features, and reports better classification than SVMs, with 93.06% modified accuracy using decision-level fusion.
Problem
Prior schizophrenia classifiers often used single-channel EEG features or a single connectivity measure, overlooking inter-channel network interactions and complementary connectivity domains.
Method
The MDC-CNN combines VAR, PDC, and complex-network connectivity features through parallel 1D- and 2D-CNNs with fusion strategies.
Results
93.06% modified accuracy was achieved with MDC-CNN decision-level fusion, which outperformed SVMs and single-domain CNNs.
Takeaways & Limitations
Integrating diverse EEG brain-connectivity descriptors enables accurate discrimination of schizophrenia from healthy controls.
Abstract
from arXiv · showhide
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.
1 Introduction
The paper addresses schizophrenia classification by modeling disrupted EEG brain connectivity as a network rather than relying on isolated regional or single-channel features. It proposes a multi-domain CNN framework to combine directed connectivity and network-topology measures for discriminating schizophrenia from healthy controls.
- 1 Introduction: Single-region abnormalities cannot fully explain schizophrenia impairments, motivating analysis of interactions across brain regions.The introduction connects observed structural, functional, and electrophysiological abnormalities with the need for network-level analysis.
- 1 Introduction: Schizophrenia is characterized as a dysconnectivity disorder involving abnormal structural and functional brain connectivity networks.Connectivity alterations are presented as potentially useful biomarkers for clinical applications.
- 1 Introduction: Prior EEG classifiers mainly used single-channel time-frequency features, which ignore inter-channel interactions that can help distinguish schizophrenia from healthy controls.Traditional approaches included SVM, KDA, and logistic regression, while network connectivity provides group-level discriminative information.
- 1 Introduction: The study proposes deep CNNs for classifying altered EEG-derived brain connectivity patterns in schizophrenia.The framework targets discriminative analysis of schizophrenia versus healthy controls using connectome features.
- 1 Introduction: The MDC-CNN combines time-domain VAR, frequency-domain PDC, and complex-network measures using parallel 1D- and 2D-CNNs.The architecture is designed to fuse complementary connectivity domains and dimensionalities.
2 Methods
The method extracts directed connectivity and network-topology features from EEG, processes them with parallel CNNs, and fuses learned representations or predictions across domains for SZ–HC classification.
- 2.1 Measures of Directed Connectivity: EEG-derived directed connectivity is characterized using VAR coefficients, PDC, and topology-based complex-network measures.VAR quantifies directed connectivity across time lags, PDC measures direct frequency-domain dependencies, and complex-network analysis summarizes higher-order organization.
- 2.1 Measures of Directed Connectivity: VAR coefficient matrices quantify directed channel influences at time lags, while PDC normalizes frequency-specific information outflow between channels.The VAR representation uses coefficient matrices across lags; PDC is defined from the Fourier transform of those coefficients.
- 2.3 Single-Domain CNN: The CNN treats connectivity matrices as image-like inputs, using time lags or frequency bands as channels for learning hierarchical spatial representations.Convolutional and pooling layers extract progressively higher-level features from the connectivity structures.
- 2.4 Proposed MDC-CNN: The MDC-CNN combines 2D VAR and PDC connectivity matrices with 1D complex-network measures through parallel 2D- and 1D-CNNs.VAR and PDC inputs encode time- and frequency-domain connectivity, while complex-network measures represent whole-network topology.
- 2.4 Proposed MDC-CNN: The framework supports feature-level, score-level, and decision-level fusion of domain-specific CNN outputs.Feature-level fusion concatenates learned maps, score-level fusion combines softmax probabilities, and decision-level fusion uses majority voting.
3 Experimental Results
The experiments evaluate EEG-connectivity CNNs against SVMs, across connectivity domains, frequency bands, fusion strategies, and learned representations. The multi-domain CNN achieves the strongest reported classification performance while requiring only 1.2 ms per subject.
- Model Architecture and Training: The 2D-CNN processes 16×16×5 VAR and PDC matrices, while the 1D-CNN learns interactions among complex-network features.Pooling was omitted from the 2D-CNN because it reduced classification accuracy.
- Model Architecture and Training: Feature maps learned from PDC matrices showed different patterns between SZ and HC, with stronger SZ activations for some first-layer filters.The model was selected at the epoch with minimum validation cross-entropy loss, which occurred at epoch 80.
- Classification Results: 92.87% accuracy and 93.06% modified accuracy were achieved by majority-voting MDC-CNN, outperforming single-domain CNNs and SVMs.The comparison used VAR, PDC, complex-network, and fused connectivity features.
- Classification Results: CNNs outperformed SVMs across all tested connectivity feature sets, while VAR features performed better than PDC and complex-network features alone.Complex-network features alone reached 84.49% accuracy with CNN and 78.60% with SVM.
- Frequency-Band Analysis: PDC-based CNN accuracy ranged from 84.49% to 88.01%, with lower EEG frequency bands more discriminative than higher bands.PDC modified accuracy ranged from 84.48% to 87.56%.
- Frequency-Band Analysis: Alpha-band complex-network features slightly exceeded PDC accuracy, whereas complex-network accuracy ranged from 71.25% to 86.99% across bands.The corresponding modified-accuracy range for complex-network features was 69.31% to 86.67%.
- Classification Results: Almost 5% higher accuracy than the best SVM was achieved by MDC-CNN, with classification taking only 1.2 ms per subject.CNN required more computation than SVM, but remained in the millisecond range.
- Fusion Analysis: Majority-voting decision-level fusion outperformed the other tested fusion strategies despite slightly higher classification cost.The fusion strategies were evaluated for the MDC-CNN.
4 Conclusion
The study develops an EEG connectome CNN framework that captures disrupted connectivity patterns in schizophrenia and combines complementary connectivity descriptors for classification.
- The MDC-CNN combines time- and frequency-domain effective-connectivity metrics with complex network measures to integrate complementary EEG connectome information.
- The CNN learns hierarchical low- and high-level representations from EEG connectivity features that differentiate schizophrenia patients from healthy controls.
- The proposed CNN outperforms traditional SVM classifiers on a large resting-state EEG dataset, supporting automated group-level discrimination.