Abstract
Resting-state functional Magnetic Resonance Imaging (rs-fMRI) using sliding-window-based dynamic functional connectivity networks (SW-D-FCN) has become a vital tool for diagnosing neurodegenerative diseases. However, traditional feature selection methods often fail to capture the personalized discriminative characteristics within individual functional connectivity (FC) sequences. This paper proposes a Unit-based Personalized Fingerprint Feature Selection (UPFFS) strategy. By treating each ROI pair's FC sequence as a unique "unit" and identifying its most discriminative feature, UPFFS creates a comprehensive "fingerprint" for classification. Evaluated on Autism Spectrum Disorder (ASD) diagnosis using the ABIDE dataset, the strategy demonstrates superior performance over conventional single-view and concatenation methods.
Introduction
The study of brain dynamics through rs-fMRI has shifted from static snapshots to Dynamic Functional Connectivity (DFC), which tracks neural interactions over time. While this provides a richer dataset, it introduces significant complexity. For every pair of brain regions, a sequence of connectivity values is generated. Extracting the most relevant features from these sequences is critical for accurate diagnosis.
Traditional approaches typically employ one of two "reductionist" strategies: Single-View Feature Selection (SVFS), which uses a single feature type (like mean) across all connections, or Concatenation Feature Selection (CFS), which merges multiple feature types into a large vector. These methods often overlook the heterogeneity of brain connections, where some may be best characterized by their average strength and others by their temporal stability.
Methodology
The proposed UPFFS strategy addresses these limitations by defining each ROI-to-ROI connection as a "unit." Instead of a global feature selection, UPFFS evaluates multiple feature types (based on central moments) for each specific unit to find its unique "fingerprint"—the feature that best distinguishes the target disease for that specific connection.
As shown in the framework below, the process involves:
- D-FCN Construction: Generating dynamic networks from rs-fMRI data using sliding windows.
- Feature Extraction: Calculating multiple feature sets (e.g., mean, variance, skewness) for every connection.
- Personalized Selection: Identifying the most discriminative feature for each unit independently.
- Classification: Concatenating these fingerprints into a final feature set for a Linear Support Vector Machine (SVM).
Results and Discussion
The UPFFS strategy was evaluated using the ABIDE (Autism Brain Imaging Data Exchange) database to classify ASD vs. Typical Controls. The results highlight several key advantages:
- Superior Accuracy: UPFFS achieved 73.6% accuracy in Low-Order networks and 74.0% in High-Order networks, consistently outperforming both SVFS and CFS.
- The Power of Fusion: By fusing the results of both Low- and High-Order networks using the UPFFS strategy, the researchers reached a peak accuracy of 76.0%.
- Balanced Selection: By ensuring every relevant connection is represented by its best feature, UPFFS avoids the selection bias common in concatenation methods, where a few noisy connections might dominate the feature set.
- Biomarker Identification: The model identified key discriminative regions—including the superior frontal gyrus, hippocampus, and calcarine cortex—that align with known social and visual processing deficits in ASD.
Conclusion
UPFFS offers a more nuanced approach to brain network analysis by respecting the local diversity of functional connections. While performance is sensitive to the choice of feature evaluation metrics, the "unit-based" logic provides a robust framework for identifying personalized biomarkers. Future work may extend this logic to other complex biological datasets, such as genomics or proteomics, where identifying personalized fingerprints within complex sequences is equally critical.
Summary updated