BACE: Behavior-Adaptive Connectivity Estimation for Interpretable Graphs of Neural Dynamics

TL;DR

BACE addresses the need for predictive yet interpretable models of dynamic brain connectivity by learning phase-specific, directed, region-level adjacency matrices from multi-region intracranial LFPs. The framework uses per-region encoders, a context-specific graph learner, and a graph projector to fuse inter-regional influences into forecasting, tying connectivity directly to dynamics via an end-to-end objective. In synthetic tests, BACE accurately recovers ground-truth directed interactions; in human LFP data, it yields explicit, phase-specific graphs that reveal behavior-aligned network reconfiguration and align with anatomical and motor organization principles. Together, these results establish BACE as a practical tool for generating data-driven hypotheses about how subcortical regions coordinate during behavior, with interpretable connectivity objects suitable for neuroscientific analysis and potential clinical applications.

Abstract

Understanding how distributed brain regions coordinate to produce behavior requires models that are both predictive and interpretable. We introduce Behavior-Adaptive Connectivity Estimation (BACE), an end-to-end framework that learns phase-specific, directed inter-regional connectivity directly from multi-region intracranial local field potentials (LFP). BACE aggregates many micro-contacts within each anatomical region via per-region temporal encoders, applies a learnable adjacency specific to each behavioral phase, and is trained on a forecasting objective. On synthetic multivariate time series with known graphs, BACE accurately recovers ground-truth directed interactions while achieving forecasting performance comparable to state-of-the-art baselines. Applied to human subcortical LFP recorded simultaneously from eight regions during a cued reaching task, BACE yields an explicit connectivity matrix for each within-trial behavioral phase. The resulting behavioral phase-specific graphs reveal behavior-aligned reconfiguration of inter-regional influence and provide compact, interpretable adjacency matrices for comparing network organization across behavioral phases. By linking predictive success to explicit connectivity estimates, BACE offers a practical tool for generating data-driven hypotheses about the dynamic coordination of subcortical regions during behavior.

Figures (4)

• Figure 1: BACE architecture. Per-region temporal encoders (independent GRUs) extract local dynamics from raw neural windows. A behavioral-specific graph learner produces one directed adjacency $\mathbf{A}_\phi$ for each behavioral segment, which the graph projector applies to mix regional features. The forecaster head then predicts the unobserved future window using a residual objective. This end-to-end design links forecasting directly to explicit, behavioral-specific connectivity estimates.
• Figure 2: Adjacency recovery on synthetic datasets. (A) Stochastic non-Gaussian suite; (B) structured suite. Columns correspond to the four datasets $\mathcal{D}_1$–$\mathcal{D}_4$. Rows: (top) ground-truth adjacency $\mathbf{A}$; (middle) BACE’s learned adjacency $\hat{\mathbf{A}}$ (row-normalized at inference); (bottom) top-$k$ binarization of $\hat{\mathbf{A}}$ with $k$ equal to the ground-truth row degree (here $k{=}2$). Numbers above each panel indicate F1@$k_{\text{row}}$. Quantitatively, BACE achieves mean correlation $0.89$ and mean F1@$k_{\text{row}}{=}1.00$ on the structured suite, and $0.90$ / $0.98$ on the stochastic suite.
• Figure 3: (A) Task schematic and behavioral segmentation timeline. Each trial comprised four stages with a 400 ms analysis window: Wait ($-400$ ms to Go cue onset), React (0–400 ms after Go cue onset), Reach (0–400 ms after forward reach onset), and Return (0–400 ms after return onset). (B) behavioral-specific $8\times8$ effective connectivity matrices estimated by BACE across eight subcortical regions (GPi1, GPi2, VIM, STN; left and right hemispheres). Matrix entry $A_\phi[i,j]$ denotes directed influence from source region $j$ (column) to target region $i$ (row).
• Figure 4: Group-level analyses of behaviorally-specific connectivity. (A) Edges showing significant differences across behavioral conditions, based on bootstrap CIs and Benjamini–Hochberg false discovery rate (FDR) correction. (B) Average outgoing connectivity strength from left vs. right hemisphere sources. (C) Average ipsilateral vs. contralateral connectivity strength.