GEFF: Graph embedding for functional fingerprinting

Kausar Abbas, Enrico Amico, Diana Otero Svaldi, Uttara Tipnis, Duy Anh Duong-Tran, Mintao Liu, Meenusree Rajapandian, Jaroslaw Harezlak, Beau M. Ances, Joaquín GoñiNeuroImage 2020 · DOI

TL;DR

This paper introduces GEFF, a novel framework that uses graph embeddings to create unique "fingerprints" of individual brains from fMRI data. GEFF significantly improves the accuracy of identifying individuals and the tasks they are performing, even across different tasks, paving the way for more personalized neuroscience and clinical applications.

What Problem Does This Paper Address?

Functional Connectomes (FCs), derived from fMRI data, have shown promise as individual "fingerprints" that can identify a person from a group. While resting-state FCs yield high identification rates, task-based FCs often show moderate to low accuracy, and identifying individuals across different tasks remains a significant challenge. This "task-dependency" limits the potential for a universal brain fingerprint. Furthermore, existing methods often struggle to distinguish between individual-specific and task-specific brain activity, hindering the development of personalized biomarkers for neurological and psychiatric disorders. The authors aim to develop a framework that can achieve high subject and task identification rates, potentially in a task-independent manner, while preserving individual variability.

Methods

The core of GEFF lies in its two-stage approach:

  1. Learning Stage: Eigenspace Embedding: GEFF first creates a low-dimensional "eigenspace" representation of functional connectomes from a group of subjects. This is achieved using Principal Component Analysis (PCA) on vectorized FCs. The resulting eigenvectors capture the main sources of variance in the data.
  2. Identification Stage: Nearest Centroid Classifier: In this stage, GEFF computes average representations (centroids) for each class (subject or task) within the learned eigenspace. A new, unseen FC is then projected into this eigenspace, and its identity (subject or task) is determined by finding the closest centroid. This approach is inspired by object recognition techniques.

A key innovation is the use of group-level decomposition to create the eigenspace, which allows for potentially task-independent identification. Unlike some previous methods, GEFF does not require test-retest data for the same subject.

Datasets

The study utilizes data from the Human Connectome Project (HCP) 900 subjects data release. This includes:

  • 100 unrelated subjects (54 females, 46 males, mean age 29.1 ± 3.7 years).
  • Eight different fMRI tasks: resting-state (RS), emotion (EM), gambling (GAM), language (LAN), motor (MOT), relational (REL), social (SOC), and working memory (WM).
  • Resting-state fMRI scans were acquired on two different days, with two sessions each. Task fMRI data also included multiple runs.

Key Insights and Ideas

  • Task-Independent Fingerprinting is Possible: GEFF demonstrates that by using a combination of resting-state and just one other task in the learning stage, it's possible to achieve high subject identification rates across a wide range of cognitive tasks. This suggests that resting-state and one other task might cover a significant portion of the "cognitive space" for individual fingerprinting.
  • Eigenvectors Reveal Structure: The study found that eigenvectors derived from the learning stage naturally fall into distinct "regimes": task-dominant (early eigenvectors explaining most variance) and subject-dominant (later eigenvectors). This provides a novel insight into how brain connectivity varies across individuals and tasks.
  • Averaging Improves Fingerprints: Creating an average individual representation using multiple task FCs, especially when combined with resting-state, significantly enhances the robustness and generalizability of the individual fingerprint.
  • Resting-State is Key for Generalization: Resting-state connectivity appears to capture a fingerprint that is somewhat orthogonal to other tasks, making it crucial for achieving broad task-independent identification.

Key Results

  • High Subject Identification (SI) Rates: GEFF significantly improved SI rates over traditional methods (Orig FCs) across all tasks and scenarios. For example, with only one learning task, GEFF achieved up to 90% SI accuracy for most tasks, and 100% for resting-state. With two learning tasks (including resting-state), GEFF achieved near-perfect SI rates (98%).
  • Task Identification (TI) Rates: GEFF achieved high TI rates, saturating around 99% with increasing subjects per task. Even with only 20 subjects per task, GEFF achieved 94.8% average TI accuracy across all eight tasks, comparable to deep learning methods.
  • Task-Independence: GEFF showed strong evidence for task-independent SI when resting-state was included in the learning set, achieving high accuracy even for tasks not used in learning.
  • Subject-Independence: GEFF also demonstrated subject-independent TI, meaning the learned task representations were generalizable to new subjects.
  • Robustness to Sample Size: GEFF's performance remained stable with increasing sample size, unlike traditional methods which showed decreased accuracy.

Limitations and Future Work

  • Sample Size: GEFF was tested on a relatively modest sample size of 100 subjects. The authors suggest testing with larger datasets to further confirm robustness.
  • Parcellation Dependency: The study used one specific brain parcellation. It's acknowledged that parcellation size can affect FC fingerprinting, and exploring different parcellations is a future direction.
  • Dynamic Updates: New FCs cannot be added dynamically to GEFF without recomputing the eigenspace. This requires a group-level decomposition for every new FC.
  • Contribution of Individual Regions: GEFF embeds whole FCs, making it difficult to discern the contribution of individual brain regions to identification accuracy.
  • Limited Task Space: The study explored cognitive space through eight fMRI tasks. Datasets with more tasks would allow for a richer exploration of subject and task fingerprints.

Future Work: * Reproduce findings with larger sample sizes and more tasks. * Explore the use of dynamic FCs to create dynamic eigenspace profiles for tracking changing mental states. * Investigate the framework with different parcellation sizes. * Apply GEFF to effective connectivity data. * Track disease progression over time using GEFF.

Summary updated

BibTeX
@article{abbas2020geff,
  author = {Kausar Abbas and Enrico Amico and Diana Otero Svaldi and Uttara Tipnis and Duy Anh Duong-Tran and Mintao Liu and Meenusree Rajapandian and Jaroslaw Harezlak and Beau M. Ances and Joaquín Goñi},
  title = {{GEFF: Graph embedding for functional fingerprinting}},
  journal = {NeuroImage},
  year = {2020},
  doi = {10.1016/j.neuroimage.2020.117181},
}
Figure