Latent-Space Causal Discovery from Indirect Neuroimaging Observations

Researchers introduce INCAMA, a novel method for inferring causal brain networks from indirect neuroimaging data like fMRI. The approach addresses the fundamental challenge that brain imaging signals are distorted by physics of hemodynamics and volume conduction, making direct causal inference impossible without accounting for these measurement artifacts.

This research tackles a longstanding neuroscience problem: neuroimaging measurements don't directly observe neural activity but rather indirect signals filtered through biological and physical processes. Standard causal inference methods fail when applied to fMRI or similar modalities because they assume measurements reflect true underlying variables. INCAMA bridges this gap by formally characterizing when causal structure can be recovered from indirect observations under explicit physical assumptions about how neural signals transform into measurable hemodynamic responses.

The method combines two key innovations: physics-aware signal inversion that accounts for hemodynamic distortion, and a Mamba encoder that exploits nonstationary dynamics and mechanism shifts to identify directed relationships. Testing on The Human Connectome Project motor-task fMRI data shows the approach produces sparse networks concentrated in known visuo-motor pathways, suggesting real-world validity beyond controlled simulations.

For the broader neuroscience community, this work establishes theoretical foundations for causal brain imaging analysis that were previously lacking. It enables researchers to make stronger claims about directed connectivity from standard neuroimaging protocols without new data collection or invasive procedures. The 2-3x improvement in directed-structure recovery over existing methods suggests practical advantages for clinical applications like understanding neural dysfunction in neurological disorders.

Future work likely focuses on scaling to whole-brain analyses, incorporating multi-modal data, and validating against invasive recordings where ground truth is available. This represents progress toward more principled causal inference in neuroscience, though limitations remain in temporal resolution and spatial specificity of fMRI.

• →INCAMA achieves 2-3x improvement in directed graph recovery over standard observation-space methods on controlled neuroimaging simulations.
• →The method formally addresses how hemodynamic distortion and volume conduction prevent direct causal inference from fMRI signals.
• →Validated on Human Connectome Project data, producing sparse networks concentrated in known anatomical motor pathways.
• →Physics-aware signal inversion coupled with delay-aware Mamba encoding enables mechanism shifts to inform causal scoring.
• →Establishes theoretical framework for recovering latent causal structure from indirect neuroimaging observations under explicit assumptions.