Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Deep Spatio-Temporal Architecture for Dynamic Effective Connectivity Network Analysis Based on Dynamic Causal Discovery

Faming Xu, Yiding Wang, Chen Qiao, Gang Qu, Vince D. Calhoun, Julia M. Stephen, Tony W. Wilson, Yu-Ping Wang

TL;DR

This work tackles the challenge of uncovering dynamic directed brain interactions by introducing STDCDAE, a deep spatio-temporal auto-encoder that fuses spatial graphs and temporal sequences to learn dynamic causal masks. The framework couples a Dynamic Causal Deep Encoder (DCDE) with a Dynamic Causal Deep Decoder (DCDD), and employs Spatial Information Fusion (SIF) and Temporal Information Embedding (TIE) to capture evolving causality, with reconstruction, structure, divergence, and sparsity losses guiding learning. Empirical results on linear VAR and nonlinear Lorenz-96 simulations show superior causal discovery performance, while resting-state fMRI data from the Philadelphia Neurodevelopmental Cohort reveal developmental shifts in dynamic effective connectivity networks, from dispersed, redundant patterns in children to more stable, specialized networks in young adults. The findings provide a data-driven view of brain maturation, highlighting how dynamic dECNs mediate information flow across Resting-State Networks (RSNs) and underpin higher cognitive abilities in adulthood, with STDCDAE offering a robust, interpretable tool for neuroscience research and potential clinical applications.

Abstract

Dynamic effective connectivity networks (dECNs) reveal the changing directed brain activity and the dynamic causal influences among brain regions, which facilitate the identification of individual differences and enhance the understanding of human brain. Although the existing causal discovery methods have shown promising results in effective connectivity network analysis, they often overlook the dynamics of causality, in addition to the incorporation of spatio-temporal information in brain activity data. To address these issues, we propose a deep spatio-temporal fusion architecture, which employs a dynamic causal deep encoder to incorporate spatio-temporal information into dynamic causality modeling, and a dynamic causal deep decoder to verify the discovered causality. The effectiveness of the proposed method is first illustrated with simulated data. Then, experimental results from Philadelphia Neurodevelopmental Cohort (PNC) demonstrate the superiority of the proposed method in inferring dECNs, which reveal the dynamic evolution of directed flow between brain regions. The analysis shows the difference of dECNs between young adults and children. Specifically, the directed brain functional networks transit from fluctuating undifferentiated systems to more stable specialized networks as one grows. This observation provides further evidence on the modularization and adaptation of brain networks during development, leading to higher cognitive abilities observed in young adults.

A Deep Spatio-Temporal Architecture for Dynamic Effective Connectivity Network Analysis Based on Dynamic Causal Discovery

TL;DR

This work tackles the challenge of uncovering dynamic directed brain interactions by introducing STDCDAE, a deep spatio-temporal auto-encoder that fuses spatial graphs and temporal sequences to learn dynamic causal masks. The framework couples a Dynamic Causal Deep Encoder (DCDE) with a Dynamic Causal Deep Decoder (DCDD), and employs Spatial Information Fusion (SIF) and Temporal Information Embedding (TIE) to capture evolving causality, with reconstruction, structure, divergence, and sparsity losses guiding learning. Empirical results on linear VAR and nonlinear Lorenz-96 simulations show superior causal discovery performance, while resting-state fMRI data from the Philadelphia Neurodevelopmental Cohort reveal developmental shifts in dynamic effective connectivity networks, from dispersed, redundant patterns in children to more stable, specialized networks in young adults. The findings provide a data-driven view of brain maturation, highlighting how dynamic dECNs mediate information flow across Resting-State Networks (RSNs) and underpin higher cognitive abilities in adulthood, with STDCDAE offering a robust, interpretable tool for neuroscience research and potential clinical applications.

Abstract

Dynamic effective connectivity networks (dECNs) reveal the changing directed brain activity and the dynamic causal influences among brain regions, which facilitate the identification of individual differences and enhance the understanding of human brain. Although the existing causal discovery methods have shown promising results in effective connectivity network analysis, they often overlook the dynamics of causality, in addition to the incorporation of spatio-temporal information in brain activity data. To address these issues, we propose a deep spatio-temporal fusion architecture, which employs a dynamic causal deep encoder to incorporate spatio-temporal information into dynamic causality modeling, and a dynamic causal deep decoder to verify the discovered causality. The effectiveness of the proposed method is first illustrated with simulated data. Then, experimental results from Philadelphia Neurodevelopmental Cohort (PNC) demonstrate the superiority of the proposed method in inferring dECNs, which reveal the dynamic evolution of directed flow between brain regions. The analysis shows the difference of dECNs between young adults and children. Specifically, the directed brain functional networks transit from fluctuating undifferentiated systems to more stable specialized networks as one grows. This observation provides further evidence on the modularization and adaptation of brain networks during development, leading to higher cognitive abilities observed in young adults.

Paper Structure

This paper contains 18 sections, 20 equations, 5 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. The Architecture of SIF and TIE
  4. Spatial Information Fusion (SIF)
  5. Temporal Information Embedding (TIE)
  6. STDCDAE
  7. Dynamic Causal Deep Encoder (DCDE)
  8. Dynamic Causal Deep Decoder (DCDD)
  9. The Learning Criteria of STDCDAE
  10. Simulation Experiments
  11. VAR Experiments
  12. Lorenz-96 Experiments
  13. Analysis of Dynamic Effective Connectivity Network in Brain
  14. Data Preprocessing and Model Hyperparameters
  15. Inference of Brain dECN
  16. ...and 3 more sections

Figures (5)

  • Figure 1: (a) The flow chart of STDCDAE, which demonstrates how the $i$-th deep auto-encoder in STDCDAE infers intrinsic dynamic causality for the $i$-th node. DCDE represents the dynamic causal deep encoder . DCDD represents the dynamic causal deep decoder. TIE and SIF represent temporal information extraction and spatial information fusion, respectively. MMG represents mask matrix generation. RL means representation learning. TIP represents temporal information processing. (b) The calculation process of the TIE module in STDCDAE. (c) The calculation process of the SIF module in STDCDAE.
  • Figure 2: The brain functional networks of children and young adults inferred by different methods, where the first row shows the ECNs inferred by different causal discovery algorithms, and the second row shows the FCNs inferred by different correlation coefficients. Compared with cMLP and cLSTM, STDCDAE reveals clearer different ECN patterns between groups. Meanwhile, compared with the FCNs revealed by correlation coefficients, the ECNs inferred by STDCDAE not only capture the directed relationship among ROIs, but also exhibit higher sparsity, which is helpful for revealing the essential differences between different groups.
  • Figure 3: (a) The proportion of different connections between children and young adults. (b) The distribution of dECs in different RSNs and the information interaction patterns of each RSN in children. (c) The distribution of dECs in different RSNs and the information interaction patterns of each RSN in young adults. (d) The distribution of enhanced dECs in different RSNs. (e) The distribution of weakened dECs in different RSNs.
  • Figure 4: (a) The fluctuation of dEC intensity over time between ROIs and the fluctuation of dEC intensity within and between RSNs over time in children. (b) The fluctuation of dEC intensity over time between ROIs and the fluctuation of dEC intensity within and between RSNs over time in young adults.
  • Figure 5: The changes in information exchange patterns of brain functional network hubs over time. In this figure, the x-axis represents scan time and the y-axis represents the information flow intensity, every point stands for the directed flow intensity between RSNs of a subject at that time. (a) Changes in SSN information interaction patterns for children and young adults over time, with blue representing children and yellow representing young adults (the same below). The mean and 95% confidence interval are shown in (d). (b) Changes in DMN information interaction patterns for children and young adults over time. The mean and 95% confidence interval are shown in (e). (c) Changes in VN information interaction patterns for children and young adults over time. The mean and 95% confidence interval are shown in (f).