Machine Learning on Dynamic Functional Connectivity: Promise, Pitfalls, and Interpretations

TL;DR

This study benchmarks a wide range of deep learning approaches for dynamic functional connectivity derived from fMRI, across task-evoked and resting-state paradigms, using 34,887 samples from six public datasets. It contrasts spatial graph models with sequential time-series models, finding that sequential architectures excel in task fMRI while spatial models perform better on resting-state disease diagnosis, with SPDNet and Transformer leading in their respective domains. The authors also explore post-hoc brain mappings to assess model explainability, revealing both biologically plausible regions and gaps due to representation misalignment. They propose neuroscience-informed guidelines for backbone selection and emphasize explainability to facilitate clinical translation, while releasing preprocessed benchmark data for reproducible research.

Abstract

An unprecedented amount of existing functional Magnetic Resonance Imaging (fMRI) data provides a new opportunity to understand the relationship between functional fluctuation and human cognition/behavior using a data-driven approach. To that end, tremendous efforts have been made in machine learning to predict cognitive states from evolving volumetric images of blood-oxygen-level-dependent (BOLD) signals. Due to the complex nature of brain function, however, the evaluation on learning performance and discoveries are not often consistent across current state-of-the-arts (SOTA). By capitalizing on large-scale existing neuroimaging data (34,887 data samples from six public databases), we seek to establish a well-founded empirical guideline for designing deep models for functional neuroimages by linking the methodology underpinning with knowledge from the neuroscience domain. Specifically, we put the spotlight on (1) What is the current SOTA performance in cognitive task recognition and disease diagnosis using fMRI? (2) What are the limitations of current deep models? and (3) What is the general guideline for selecting the suitable machine learning backbone for new neuroimaging applications? We have conducted a comprehensive evaluation and statistical analysis, in various settings, to answer the above outstanding questions.

Figures (8)

• Figure 1: Various data representations in fMRI studies (in time series or graph) and machine learning models for fMRI analyses.
• Figure 2: Selected deep models for the benchmark, which include spatial (left) and sequential models (right).
• Figure 3: Significant tests between spatial models () and sequential models () for various datasets. The first and third plots for the HCP dataset are "Separated Scan 1 & Scan 2" experiment (marked as '###-1'), and the second and fourth plots are "Mixed Scan 1 & Scan 2 setting" (marked as '###-2').
• Figure 4: SPDNet exhibits significant performance gain over all other spatial models except for GCN.
• Figure 5: Brain mapping of logistic regression weights for features derived from different models. Black and red circles denote Motor and Language-related brain areas, respectively.