TL;DR
This paper introduces a novel explainable clustering approach to understand how functional brain network connectivity (FNC) relates to cognitive differences in healthy individuals. By combining clustering with explainability methods like G2PC and logistic regression, they identified two distinct cognitive groups and pinpointed specific brain network pairs that drive these differences, offering a deeper insight into the neural underpinnings of cognition.
What Problem Does This Paper Address?
Functional Magnetic Resonance Imaging (fMRI) allows researchers to study brain activity by measuring blood flow. Functional Network Connectivity (FNC) quantifies how different brain regions or networks interact during rest. Clustering algorithms are often applied to FNC data to identify subgroups of individuals with similar brain activity patterns or to detect different "brain states." While these clusters can reveal interesting groupings, a significant challenge has been understanding why these clusters form – specifically, which FNC patterns are most influential in assigning individuals to a particular group. This lack of explainability limits our ability to translate these findings into a deeper understanding of how brain network dynamics relate to cognitive functions and disorders. This paper aims to bridge this gap by developing and applying explainable methods to FNC clustering.
Methods
The core of this study involves applying k-means clustering to static FNC (sFNC) data derived from resting-state fMRI. The novelty lies in the subsequent application of explainability approaches to interpret the clustering results.
- Clustering: K-means was used to partition the sFNC data into groups. The optimal number of clusters was determined to be two (k=2) using Elbow and Silhouette methods.
- Explainability Approaches:
- G2PC Feature Importance: This is a key contribution. The authors use a variation of Global Permutation Percent Change (G2PC), an algorithm-agnostic method that measures how sensitive clusters are to perturbations of specific features (in this case, brain network pairs). They introduce a normalized version (G2PCn) that accounts for the number of features within a network pair, providing a more nuanced view of importance compared to the unnormalized G2PCu.
- Interpretable Logistic Regression with Elastic Net Regularization (LR-ENR): This method was used as a comparative approach. By training an LR-ENR classifier, the researchers could examine the weights associated with each network pair, providing another perspective on their influence on the clustering. They further calculated the "mean effect" of each network pair by multiplying model weights with feature values.
Datasets
The study utilized data from the UK BioBank. This dataset comprises:
- Resting-state fMRI (rs-fMRI) data.
- Cognitive performance scores from various tasks.
- Demographic information (including sex).
The dataset included 37,784 healthy individuals. The rs-fMRI data was preprocessed, and 53 independent components (ICs) associated with 7 brain networks were extracted. These networks were then used to calculate static functional network connectivity (sFNC) features, resulting in 1,378 features representing connections between pairs of ICs, which were further grouped into 28 network pairs for analysis.
Key Insights and Ideas
- Two Distinct Cognitive Profiles Emerge: The clustering successfully identified two groups of individuals with statistically significant differences in cognitive performance, suggesting that even within a "healthy" population, there are distinct underlying brain connectivity profiles related to cognition.
- The Cerebellar Network (CBN) is Crucial: Both explainability methods consistently highlighted the Cerebellar Network (CBN) and its connections (e.g., CBN/SCN, CBN/SMN) as having the most significant influence on the clustering. This suggests the CBN plays a pivotal role in differentiating these cognitive groups.
- Magnitude vs. Network Size Matters: Comparing G2PCu (unnormalized) and G2PCn (normalized) revealed important nuances. G2PCu showed the influence of larger networks (more features), while G2PCn, by accounting for feature count, provided a clearer picture of the intrinsic importance of specific network pairs, regardless of their size. This highlights the importance of careful methodological choices in explainability.
- Cognitive Differences Linked to Specific Networks: The study found that individuals in the smaller group (G0) generally performed better on visuo-spatial, reasoning, and memory tasks, and recalled digits more quickly. These cognitive advantages were linked to specific network interactions, particularly those involving the CBN and sensorimotor networks (SMN), suggesting a neural basis for these performance differences.
Key Results
- Two Clusters Identified: The k-means clustering resulted in two groups: G0 (9,967 individuals) and G1 (27,817 individuals).
- Connectivity Differences: Group G0 exhibited generally higher magnitudes of connectivity compared to G1 across many feature pairs.
- Cognitive Performance Differences: Significant differences (p < 0.05) were observed between G0 and G1 in several cognitive tests, including Numeric Memory (NM1, NM2), Fluid Intelligence (FI2), and Pair Matching (PM1, PM2, PM3). G0 generally showed better performance (e.g., remembered more digits, took less time for pair matching).
- Sex Differences: A significant difference in sex distribution was found between the groups, with G0 having a higher proportion of males, aligning with known sex differences in brain connectivity and cognitive performance.
- Top Influential Networks: The most influential network pairs identified by G2PCn and LR-ENR included CBN, CBN/SCN, SMN/SCN, CBN/SMN, and SMN.
Limitations and Future Work
- Static vs. Dynamic Connectivity: The study used static FNC (sFNC), which captures average connectivity over the entire resting-state scan. The authors acknowledge that using dynamic FNC (dFNC), which captures temporal fluctuations in connectivity, could provide more insight into dynamical patterns related to cognitive performance.
- High Dimensionality and Feature Perturbation: Due to the high dimensionality of FNC data (over 1,000 features), individual features were not sensitive enough to perturbation for G2PC to pick up their unique impact. This led to analyzing network pairs as groups, which somewhat limited the granularity of the explainability.
- Need for More Sensitive Explainability: The authors suggest that future work could benefit from more sensitive explainability approaches as they are developed, particularly for high-dimensional data.
- Statistical Relationships: A next step proposed is to directly investigate statistical relationships between the identified network pairs and specific cognitive scores that showed differences between the clusters.
Summary updated