Abstract
The wide variation in symptoms of neurological disorders among patients necessitates uncovering individual pathologies for accurate clinical diagnosis and treatment. Current methods attempt to generalize specific biomarkers to explain individual pathology, but they often lack analysis of the underlying pathogenic mechanisms, leading to biased biomarkers and unreliable diagnoses. To address this issue, we propose a motif-induced subgraph generative learning model (MSGL), which provides multi-tiered biomarkers and facilitates explainable diagnoses of neurological disorders. MSGL uncovers underlying pathogenic mechanisms by exploring representative connectivity patterns within brain networks, offering motif-level biomarkers to tackle the challenge of clinical heterogeneity. Furthermore, it utilizes motif-induced information to generate enhanced brain network subgraphs as personalized biomarkers for identifying individual pathology. Experimental results demonstrate that MSGL outperforms baseline models. The identified biomarkers align with recent neuroscientific findings, enhancing their clinical applicability.
Introduction
Neurological disorders, such as Autism Spectrum Disorder (ASD), present a significant diagnostic challenge due to their inherent clinical heterogeneity. Patients sharing the same diagnosis often exhibit vastly different symptoms and underlying brain connectivity patterns. Traditional machine learning approaches typically seek "universal" biomarkers—common brain regions or connections across a group. However, these generalized markers often fail to capture individual pathology, leading to biased results and limited explainability for clinicians.
To overcome these limitations, the Motif-induced Subgraph Generative Learning (MSGL) framework shifts the focus from simple node or edge analysis to motifs: complex, recurring high-order connectivity patterns. By identifying these motifs and using them to generate personalized brain subgraphs, MSGL provides biologically authentic biomarkers that account for individual variation while maintaining diagnostic accuracy.
Methodology
The MSGL framework is structured into three primary modules that bridge the gap between group-level patterns and individual-specific diagnosis.
1. Motif Extraction
The process begins by identifying recurring connectivity patterns across the dataset. Using a "farthest point sampling" method, the model ensures a diverse exploration of the brain's architecture. To distinguish between group-specific patterns (e.g., ASD vs. Healthy Controls), the framework employs TF-IDF (Term Frequency-Inverse Document Frequency), a technique borrowed from natural language processing, to rank and select the most pathologically relevant motifs.
2. Motif-induced Graph Generative Learning
The core innovation of MSGL is its generative approach. Rather than merely analyzing static graphs, the model generates a personalized subgraph for each patient. This process is "motif-induced," where the previously extracted motifs act as a guide, biasing the generative process toward connection patterns known to be pathologically significant. This results in a "personalized biomarker" graph tailored to each subject's unique brain structure.
3. Subgraph Representation Learning
The generated subgraphs are then processed by a Graph Neural Network (GNN). In this stage, motifs serve as "receptive fields," allowing the model to aggregate information from central brain regions and learn a final representation for classification.
Experimental Setup
The study evaluated MSGL using the ABIDE (Autism Brain Imaging Data Exchange) dataset, a prominent real-world psychiatric resource.
- Dataset Composition: 1,099 total samples (528 ASD patients and 571 healthy controls).
- Data Source: Resting-state functional Magnetic Resonance Imaging (fMRI).
- Preprocessing: The brain was parcellated into 116 Regions of Interest (ROIs) using the Automated Anatomical Labeling (AAL) template. Edges were defined based on Pearson correlation coefficients between ROI time series.
Results and Discussion
MSGL demonstrated superior performance compared to state-of-the-art baselines, including BrainGNN and BrainIB, across all primary metrics on the ABIDE dataset. The model achieved an accuracy of 71.4% (a 2.3% improvement over the best baseline) and an F1-score of 73.6%.
Beyond statistical performance, the model provided significant neuroscientific insights. It identified enhanced connectivity within the Somatomotor Network (SMN), Limbic Network (LIN), and Frontoparietal Network (FPN). Furthermore, the model highlighted specific regions such as the precuneus and temporal pole, both of which are well-documented in existing medical literature as being associated with ASD pathology. This alignment with established neuroscience underscores the model's clinical validity and explainability.
Conclusion
MSGL introduces a novel paradigm for neurological disorder detection by treating motifs as the fundamental units of analysis. By generating personalized subgraphs, the model moves beyond the search for universal biomarkers, offering a more nuanced and explainable approach to clinical diagnosis. Future research will focus on reducing the computational complexity of the autoregressive graph generation process and exploring the application of motif-based biomarkers to other heterogeneous conditions, such as schizophrenia and Alzheimer’s disease.
Summary updated