Topological inference on brain networks with application to lesion symptom mapping

1. Problem Statement

Traditional Lesion Symptom Mapping (LSM) methods (e.g., Voxel-based or Region-based LSM) treat brain areas as isolated units or focus on pairwise connections (edges) between regions. However, cognitive functions like language rely on distributed, interconnected networks where higher-order topological structures (loops, cycles) play a critical role.

• Limitations of Current Methods: Existing connectome-based LSM (CLSM) often relies on mass univariate testing of edges at a single density threshold, providing only a static snapshot. It fails to capture mesoscale homological features (e.g., 1-cycles or closed loops) that emerge and disappear across multiple scales of connectivity.
• Statistical Challenge: Comparing brain networks using Persistent Homology (PH) is difficult because Persistence Diagrams (PDs) are non-Euclidean, heterogeneous, and lack a natural vector space structure. Small data perturbations (noise) can drastically shift the "birth" and "death" coordinates of topological features, making standard parametric tests unreliable. Furthermore, existing permutation tests for PDs are often computationally prohibitive for large datasets or limited to two-group comparisons.

2. Methodology

The authors propose a Topological Lesion Symptom Mapping (TLSM) framework that integrates Persistent Homology with a novel statistical inference procedure based on Heat Kernel (HK) representation.

A. Topological Data Construction

• Filtration: Brain networks are modeled as weighted graphs. A Rips filtration is constructed by varying a threshold ($\epsilon$) on edge weights. As the threshold increases, edges are added, forming clusters and cycles.
• Persistence Diagrams (PDs): The birth and death times of topological features (specifically 1-cycles, representing closed loops of connections) are recorded as points in a PD.
• Focus: The study focuses on 1-cycles involving frontal, temporal, and parietal regions, which are critical for language processing.

B. Heat Kernel (HK) Representation

To enable statistical inference, the authors transform raw PDs into a vectorized algebraic representation:

• Smoothing: PDs are treated as heat sources (Dirac delta functions) on a triangular domain $T = \{(x,y) : y > x\}$.
• Eigenfunction Expansion: The heat diffusion equation is solved using the Laplace-Beltrami (LB) operator eigenfunctions. The PD is represented by a finite set of Fourier coefficients ($f_k$) derived from this expansion.
• Advantage: This provides a stable, vectorized representation that allows for direct comparison of PDs at the same coordinates without needing complex point-matching algorithms (like the Hungarian algorithm). It also acts as a smoothing mechanism to reduce topological noise.

C. Statistical Inference: Transposition-Based Permutation Tests

The paper introduces two novel non-parametric tests based on spectral transpositions:

1. Two-Sample Transposition Test: Compares two groups of PDs. Instead of randomizing all labels (which is computationally expensive), the method uses transpositions (swapping one subject from Group A with one from Group B).
   • Efficiency: The test statistic (L2-distance between group means in HK space) is updated incrementally in constant time ($O(1)$) per transposition, rather than recalculating from scratch.
2. Topological ANOVA (T-ANOVA): Extends the framework to compare multiple groups ($K > 2$) without dimensionality reduction.
   • It defines a ratio statistic ($\phi$) based on Topological Between-Group Sum of Squares (TSSB) and Within-Group Sum of Squares (TSSW) derived from pairwise L2-distances.
   • Like the two-sample test, it updates sums of squares incrementally during random transpositions of group labels.

3. Key Contributions

1. Novel Inference Framework: Development of a stable, heat-kernel-based statistical framework for comparing Persistence Diagrams that avoids the "point matching" bottleneck.
2. Efficient Algorithms: Introduction of spectral transposition tests that significantly reduce computational complexity compared to standard permutation tests, enabling analysis of large-scale brain network data.
3. Topological ANOVA (T-ANOVA): A new method for multi-group comparison of topological features that operates directly on the functional means of HK estimates, preserving the full topological structure without dimensionality reduction (unlike Isomap-based approaches).
4. TLSM Application: The first application of topological inference to Lesion Symptom Mapping, shifting the focus from individual edges to the robustness of mesoscale connectivity loops (1-cycles).

4. Results

A. Simulation Studies

• Point Clouds: The methods were tested on synthetic data with known topological structures (key shapes with holes). The tests demonstrated high sensitivity in detecting differences in cycle presence and robustness against topological noise and variations in cycle location.
• Brain Network Simulations: Using real resting-state fMRI data with simulated "node knockouts" (mimicking lesions):
  • Type I Error: All methods maintained excellent control (p-values $\approx$ 1) when no topological difference existed.
  • Power: The proposed Topological Transposition and T-ANOVA achieved near-zero p-values when lesions were introduced, demonstrating superior sensitivity compared to Heo's ANOVA (which often failed to detect differences) and PERMANOVA.
  • Efficiency: Computational time for T-ANOVA grew linearly/steadily with sample size, whereas competing methods showed exponential growth.

B. Real-World Application (Post-Stroke Aphasia)

• Dataset: Resting-state fMRI from 96 patients with anomic or Broca's aphasia.
• Grouping: Patients were stratified by behavioral performance (fluency, repetition, comprehension, naming) into low vs. high subscore groups.
• Findings:
  • Traditional CLSM and Network-Based Statistic (NBS) methods identified very few or no significant connections.
  • TLSM successfully identified significant differences in the persistence patterns of 4-polygons (1-cycles involving 4 nodes).
  • Specific Results: Significant differences were found in Comprehension and Naming categories.
  • Biological Insight: The study revealed that patients with lower scores had "denser" but potentially less robust loop structures, while high-persistence loops (indicating resilient, redundant pathways) were preferentially associated with better behavioral outcomes. Specifically, the presence of a single node in the left hemisphere within a loop structure was found to be a critical differentiator for higher scores.

5. Significance

• Paradigm Shift: TLSM moves beyond the "node" and "edge" paradigm of traditional neuroimaging to analyze higher-order topological structures (loops/cycles) that are biologically interpretable as redundant or alternative communication pathways.
• Clinical Utility: The method provides a more complete picture of post-stroke deficits by linking the robustness of integrative circuits to language impairment. It suggests that recovery may depend on the preservation of these specific topological loops rather than just the integrity of individual connections.
• Methodological Impact: The framework offers a general, non-parametric, and computationally efficient tool for statistical inference on complex, non-Euclidean data structures, with potential applications beyond neuroscience (e.g., clustering, deep learning data augmentation).

In conclusion, this paper establishes a rigorous statistical pipeline for comparing brain network topology, demonstrating that topological features (specifically 1-cycles) are sensitive biomarkers for language impairment in stroke patients, offering a complementary and powerful alternative to traditional lesion mapping techniques.