Multiscale Community-Based Fingerprinting of Signed Functional Networks

• Functional connectivity patterns are subject-specific and stable across sessions and tasks, supporting individualized brain characterization.Source: §1 · p. 1
• Existing fingerprinting methods commonly use whole connectivity matrices, subnetworks, embeddings, tensors, or machine-learning features.Source: §1 · p. 1
• Most existing methods rely on univariate or edge-level features rather than multivariate descriptors of functionally coherent organization.Source: §1 · p. 1
• The proposed approach shifts fingerprinting from isolated edges to multivariate community-level topological signatures.Source: §1 · p. 2
• Signed multilayer modularity incorporates positively correlated and anti-correlated interactions when identifying subject-specific communities.Source: §1 · p. 2
• Discriminative Connectivity Maps connect identification statistics with functional neuroanatomy at ROI or functional-system levels.Source: §1 · pp. 2–3
• Community-based fingerprints are extracted from multiscale partitions, ranked by mean pairwise Euclidean distance, and used for nearest-neighbor subject identification.Source: Figure 1 · p. 3

• Functional connectivity represents pairwise co-fluctuations between brain regions as a weighted, undirected signed graph.Source: §2.1 · p. 4
• Signed-network communities favor positive connections within communities and predominantly negative connections between communities.Source: §2.2 · p. 4
• Signed modularity optimization assigns each node to exactly one community in a hard, non-overlapping partition.Source: §2.2 · p. 4
• The resolution parameter γ controls detected community size, while the signed null model separates positive and negative adjacency contributions.Source: §2.2 · p. 4
• Multilayer modularity uses interlayer coupling ω to control consistency between network layers.Source: §2.3 · p. 4
• Multiplex networks share the same nodes across layers while allowing their topological structures to differ.Source: §2.3 · p. 4
• Figure 2 reports within-task fingerprinting for S = 100, while Figure 3 compares methods at S = 200, 300, and 400.Source: Figure 2 · p. 5, Figure 3 · p. 5

3 Multiscale Community-Based Functional Connectome Fingerprinting

• Functional connectivity networks from multiple subjects are modeled as layers of a signed multilayer network.Source: §3 · p. 5
• Signed multilayer modularity generates M multiscale partitions containing joint community assignments across subjects.Source: §3 · p. 5, §3.1.1 · p. 6
• Each subject receives a low-dimensional community-based fingerprint vector for every candidate partition and graph-theoretic feature type.Source: §3 · pp. 5–6
• Selected partitions support training and testing fingerprints in a common feature space for direct subject identification.Source: §3 · p. 6
• The framework incorporates both correlated and anti-correlated activity while accommodating inter-subject variability in community assignments.Source: §3.1.1 · p. 6
• Candidate partitions arise from multiple modularity-parameter combinations, revealing joint community structures at distinct topological scales.Source: §3.1.1 · p. 7

3.2 Fingerprint Extraction

• For each candidate partition, a fingerprint component records a graph-theoretic feature computed for each community.Source: §3.2 · p. 7
• The framework evaluates participation coefficient and diversity coefficient features independently across candidate partitions.Source: §3.2 · p. 7

3.2.1 Participation coefficient

• Participation coefficient measures global inter-modular integration.Source: §3.2.1 · p. 8
• Positive and total participation coefficients use positive edges alone or the entire signed network, respectively.Source: §3.2.1 · p. 8

3.2.2 Diversity coefficient

• Diversity coefficient quantifies regional connection diversity using Shannon’s normalized entropy.Source: §3.2.2 · p. 8
• Positive and total diversity coefficients are computed from positive edges alone or from the entire signed network, respectively.Source: §3.2.2 · p. 8

3.3 Partition Selection

• Mean pairwise Euclidean distance is computed across subjects for each feature type and partition.Source: §3.3 · p. 8
• Partitions are ranked in descending order by their MPED score.Source: §3.3 · p. 9
• The top-L partitions are retained to represent complementary organizational patterns across different topological scales.Source: §3.3 · p. 9
• L is the smallest number of ranked partitions attaining at least 99% of the maximum MPED.Source: §3.3 · p. 9

3.4 Subject Identification

• Selected partitions generate training fingerprints from optimization scans and testing fingerprints from corresponding unseen scans.Source: §3.4 · p. 9
• Using the same partitions places training and testing scans in a common feature space for direct comparison.Source: §3.4 · p. 9
• Each testing fingerprint is matched against corresponding training fingerprints for its partition and feature type.Source: §3.4 · p. 9
• A distance-based nearest-neighbor classifier predicts the subject whose training fingerprint minimizes Euclidean distance to the testing fingerprint.Source: §3.4 · p. 9
• Identification success rate is the percentage of correctly identified subjects among all subjects.Source: §3.4 · p. 9

Community distinctiveness

• Self-reliability measures how stable a community-based fingerprint is across subjects and sessions using test–retest ICC.Relational similarity averages off-diagonal ICC values to capture overlap or redundancy with other communities.Source: §3.5 · p. 9
• Distinctiveness combines high self-reliability with low relational similarity to identify communities that are stable and nonredundant.Communities are ranked in descending order of their distinctiveness scores.Source: §3.5 · p. 9

Spatial interpretation

• The highest-ranked communities are selected by identifying their constituent ROIs within the joint community assignment matrix.This links community rankings to the spatial composition of the detected communities.Source: §3.5 · p. 9
• Joint community assignments support interpretation across subjects or cohorts and across individual nodes or grouped functional brain systems.The reported analysis focuses on cohort-level, system-level representations.Source: §3.5 · p. 10

Computational considerations

• The framework has expected computational complexity O(MSN^2) across M partitions for S subjects and N nodes.Signed multilayer Louvain optimization costs O(SN^2) per partition.Source: §3.6 · p. 10

Experimental design

• The evaluation covered within-task and between-task fingerprinting across eight HCP fMRI conditions, including resting state and seven tasks.The included tasks were emotion, gambling, language, motor, relational, social, and working memory processing.Source: §4 · p. 10, §4.1 · p. 10
• The study compared CBF variants with Corr-FC, GEFF, SVM, and neural-network fingerprinting methods.The proposed variants were CBF-PC+, CBF-PC, CBF-DC+, and CBF-DC.Source: §4.2 · p. 10

Small-scale evaluation

• Ten independent subject groups of S=100 were analyzed to assess robustness to subject selection across feature types and tasks.Repeating the pipeline across randomly selected groups reduced dependence on a particular subject group.Source: §4.3.1 · p. 11
• CBF-PC+ achieved the highest identification accuracy among proposed features across tasks and outperformed competing methods for most tasks, although NN was slightly better for REST1 and LAN.CBF-PC followed CBF-PC+, while CBF-DC+ and CBF-DC remained competitive.Source: §4.3.1 · p. 11

Large-scale evaluation

• As sample size increased to S=200, 300, and 400, accuracy decreased, but CBF-PC+ maintained relatively high performance across tasks and cohort sizes.Baseline methods were more sensitive to sample size and task condition, particularly on task-based fMRI.Source: §4.3.2 · p. 12
• The three-way ANOVA found significant effects of task, sample size, and method, with CBF-PC+ significantly outperforming all competing methods at adjusted p<0.05.Significant interactions also occurred among method, sample size, and fMRI task.Source: §4.3.2 · p. 12

Between-task evaluation

• CBF-PC+ produced the strongest cross-task performance at S=400, with the highest off-diagonal accuracies across nearly all task pairs.Other community-based features also outperformed competing approaches, whose cross-task generalization was lower.Source: §4.4 · pp. 13–14

Motivation and contribution

• Community-based fingerprints address the limited interpretability of edge-level methods by representing individual differences in mesoscale network organization.The approach captures functionally meaningful subgraphs rather than isolated connections.Source: §5 · p. 14

Identification performance

• The proposed method outperforms state-of-the-art approaches across both within-task and between-task identification settings.The comparison used one scan per subject under the same experimental conditions for all methods.Source: §5.1 · pp. 14–15

Task-specific organization

• Higher-order association networks, including DMN, FPN, and DAN, consistently contribute across tasks, indicating that subject variability is encoded within large-scale integrative networks.The fingerprints are neither globally uniform nor confined to a single functional network.Source: §5.2 · pp. 15–16

Robustness of spatial patterns

• Top-1% and top-5% discriminative connectivity maps show high spatial consistency across all eight tasks.Additional highly discriminative communities refine the fingerprint without fundamentally changing its task-specific organization.Source: §5.2 · p. 16

Limitations and future work

• The principal limitation is the scalability of modularity maximization for larger subject cohorts, compounded by stochastic variability in community detection.Future work proposes parallel implementations and alternative optimization strategies.Source: §5.3 · p. 16

Conclusion

• The framework represents each subject by community-based fingerprints that capture mesoscale organization across fMRI tasks and sessions.It incorporates correlations and anticorrelations to identify subject-specific communities and derives graph-theoretic fingerprints.Source: §6 · p. 16

Compliance with Ethical Standards

• The retrospective study used open-access human subject data, and the attached license indicated that ethical approval was not required.Source: §7 · p. 17