Abstract
Schizophrenia (SZ) is a debilitating functional mental disorder that significantly impacts social functioning. Accurate diagnosis is critical, yet clinical methods remain largely subjective. This study proposes an automated identification method leveraging multi-view graph measures derived from resting-state functional magnetic resonance imaging (rs-fMRI). By constructing individual functional connectivity networks based on the Brainnetome atlas and extracting diverse topological features, the researchers applied Sparse Group Lasso (SGLasso) to select discriminative features while respecting the brain's anatomical structure. Using a Support Vector Machine (SVM) classifier on a dataset of 145 subjects, the method achieved a peak accuracy of 93.10%, demonstrating superior performance over traditional single-view approaches.
Introduction
The clinical diagnosis of Schizophrenia has historically relied on psychiatric evaluations, which lack objective biological markers. While rs-fMRI offers a window into the brain's functional architecture, previous attempts to use brain network analysis for diagnosis have often failed to capture the complex topological relationships inherent in neural connectivity.
The brain is a sophisticated graph where a single region's importance can be characterized by various metrics, such as its connectivity degree, its role as a bridge (betweenness centrality), or the local efficiency of its neighbors. This paper addresses the limitations of previous models by integrating these "multi-view" graph measures. Crucially, the proposed method utilizes a feature selection process that acknowledges the anatomical grouping of these measures, ensuring that the machine learning model interprets related topological features within the same physical brain region collectively.
Methods
The researchers developed a comprehensive pipeline for SZ identification, transitioning from raw imaging data to a final classification.
- Network Construction: The brain was parcellated into 246 regions using the Brainnetome atlas. Functional connectivity networks were then constructed for each subject using Pearson correlation coefficients calculated from the time series of these regions.
- Multi-View Feature Extraction: To capture a rich "fingerprint" of brain topology, five local graph measures were calculated for every region: Degree, Betweenness Centrality, Nodal Clustering Coefficient, Local Efficiency, and Participation Coefficient. This resulted in a high-dimensional feature set of 1,230 variables per subject.
- Sparse Group Lasso (SGLasso): To manage this complexity, SGLasso was employed for feature selection. Unlike standard methods that treat variables independently, SGLasso groups the five graph measures by their corresponding brain region. This approach identifies the most discriminative brain regions while simultaneously selecting the most relevant measures within those regions, maintaining the biological relevance of the model.
- Classification: A Linear Support Vector Machine (LSVM) was trained on the optimized feature set to classify subjects as either healthy controls (HC) or patients with Schizophrenia (SZ).
Datasets
The study utilized data from the Center for Biomedical Research Excellence (COBRE). The final analysis included 145 subjects: 71 patients diagnosed with Schizophrenia and 74 healthy controls. The dataset comprised both resting-state fMRI and anatomical MRI scans.
Key Results and Insights
The proposed multi-view approach yielded significant improvements in diagnostic accuracy. The model achieved a mean classification accuracy of 93.10% using a leave-one-out cross-validation (LOOCV) scheme. The performance was highly balanced, with a sensitivity of 92.96% and a specificity of 93.24%.
The researchers found that the network's sparsity, controlled by a threshold ($t$), significantly influenced performance. The optimal results were obtained at $t=0.30$, where only the top 30% of strongest connections were retained. This accuracy significantly outperformed other feature selection methods, such as standard Lasso (83.45%) and Elastic Net (85.52%), as well as previous benchmark studies on the same dataset which typically reached 74-77% accuracy.
Furthermore, the SGLasso method identified 21 salient brain regions most discriminative for SZ, including the superior frontal gyrus, thalamus, and insular gyrus. These findings align with existing literature regarding cognitive and sensory processing impairments in Schizophrenia, suggesting that the model captures meaningful biological markers of the disorder.
Conclusion and Future Work
By treating different graph-theoretic properties as related "views" and respecting anatomical structure during feature selection, this pipeline offers a robust framework for objective SZ diagnosis. However, future work is needed to validate these findings on larger, multi-site datasets to ensure generalizability. Additionally, incorporating dynamic functional connectivity—analyzing how these graph measures change over time—could provide even deeper insights into the fluctuating nature of the disorder.
Summary updated