Causal Graph Recovery in Neuroimaging through Answer Set Programming

Mohammadsajad Abavisani, Kseniya Solovyeva, David Danks, Vince Calhoun, Sergey Plis · 2025

Read Causal Graph Recovery in Neuroimaging through Answer Set Programming by Mohammadsajad Abavisani, Kseniya Solovyeva, David Danks, Vince Calhoun, Sergey Plis (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-09286.txt.

Introduction

Mohammadsajad Abavisani and co-authors recover cause and effect graphs from brain recordings in a 2025 paper. Brain activity is sampled far slower than the real causal steps, so many graphs can fit the same data.

Their method builds the sampling loss into the search itself. A logic based solver, answer set programming, finds the best graph and also returns the family of equally possible graphs for experts to judge. Graph theory rules then prune that family, giving a smaller, sharper answer set much faster than older tools. On simulated data and on real structural brain connections, the method beat established rivals. Laid over those rivals as a second pass, it lifted their scores by about twelve percent on average.

Catalog record

CREATORMohammadsajad Abavisani, Kseniya Solovyeva, David Danks, Vince Calhoun, Sergey Plis
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE29 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2506-09286

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Causal Graph Recovery in Neuroimaging through Answer Set Programming Mohammadsajad Abavisani, Kseniya Solovyeva, David Danks, Vince Calhoun, Sergey Plis Abstract Learning graphical causal structures from time series data presents significant challenges, especially when the measurement frequency does not match the causal timescale of the system. This often leads to a set of equally possible underlying causal graphs due to information loss from sub-sampling (i.e., not observing all possible states of the system throughout time). Our research addresses this challenge by incorporating the effects of sub-sampling in the derivation of causal graphs, resulting in more accurate and intuitive outcomes. We use a constraint optimization approach, specifically answer set programming (ASP), to find the optimal set of answers. ASP not only identifies the most probable underlying graph, but also provides an equivalence class of possible graphs for expert selection. In addition, using ASP allows us to leverage graph theory to further prune the set of possible solutions, yielding a smaller, more accurate answer set significantly faster than traditional approaches. We validate our approach on both simulated data and empirical structural brain connectivity, and demonstrate its superiority over established methods in these experiments. We further show how our method can be used as a meta-approach on top of established methods to obtain, on average, 12 % 12\% improvement in F1 score. In addition, we achieved state of the art results in terms of precision and recall of reconstructing causal graph from sub-sampled time series data. Finally, our method shows robustness to varying degrees of sub-sampling on realistic simulations, whereas other methods perform worse for higher rates of sub-sampling. 1 Introduction Causal inference from functional Magnetic Resonance Imaging (fMRI) data has emerged as a critical endeavor to understand the neural mechanisms underlying cognitive processes and behaviors. Researchers not only seek to identify active brain regions during tasks, but also unravel the causal relationships between these regions, often referred to as ”effective connectivity” ( Friston 1994 ) . Graphical causal models, such as causal Bayesian networks, have become a popular framework for this purpose, combining directed graphs with joint probability distributions to model the dependencies between different brain regions ( Pearl 2009 ) . These models adhere to the Causal Markov Condition, which asserts that each node in a causal graph is conditionally independent of its non-descendants given its parents ( Spirtes et al. 2001 ) . However, applying these models to fMRI data is fraught with challenges, particularly due to the mismatch between the temporal resolution of fMRI and the rapid timescale of neural processes. The typical sampling intervals in fMRI, ranging from one to three seconds, are much slower than the millisecond-level interactions between neurons, leading to significant undersampling ( Valdes-Sosa et al. 2011 ) . This undersampling often results in the identification of multiple causal graphs that are statistically indistinguishable, forming what is known as a Markov Equivalence Class ( Pearl 2009 ; Spirtes et al. 2001 ) . These problems are exacerbated by the indirect nature of the BOLD signal, which reflects neural activity through complex and variable hemodynamic responses ( Handwerker et al. 2004 ) . In addition, the inherent variability in the hemodynamic response across different brain regions and subjects, adds another layer of complexity. Variations in the time-to-peak of the BOLD response can lead to incorrect inferences about the direction of causality, particularly when using methods like Granger causality, which assumes a fixed temporal relationship between cause and effect ( David et al. 2008 ; Seth et al. 2013 ) . Although some studies suggest that Granger causality may be robust to certain variations in the hemodynamic response ( Seth et al. 2013 ) , the combination of measurement noise, undersampling, and hemodynamic variability often undermines the reliability of causal inferences from fMRI data. Figure 1: Top left(A): Size of optimization solution set across different undersamplings, repeated 100 times. Bottom right(B): Commission error of the solution vs. optimization cost for that solution. Solutions in one equivalent class are highlighted in red. In response to these challenges, this paper proposes a novel approach that explicitly accounts for the effects of undersampling in the derivation of causal graphs. By employing constraint optimization through state-of-the-art Answer Set Programming (ASP), we aim to identify the most probable causal graph from a set of potential candidates. ASP allows for the incorporation of domain-specific knowledge and constraints, facilitating the identification of not only a single graph but an equivalence class of possible graphs, thereby offering a more comprehensive understanding of the underlying causal structure ( Gebser et al. 2012 ) . Furthermore, we introduce a method to enhance the robustness and accuracy of causal inference by integrating data from multiple instances of the same underlying causal system, such as combining fMRI data from several subjects performing the same task. This approach synthesizes information across these instances into a singular causal graph, leveraging ASP’s ability to handle complex combinatorial problems that would otherwise be infeasible with conventional techniques ( Lifschitz et al. 2008 ) . We validate our approach using both simulated data and real fMRI data from macaque brains, demonstrating its superiority over existing methods ( Smith et al. 2011 ; Ramsey et al. 2010 ) . Our results suggest that ASP offers a powerful new tool for causal inference in neuroimaging, providing more accurate and intuitive insights into the brain’s functional architecture. Figure 2: Normalized omission (top) and commission (bottom) errors for edge-breaking experiments with varying undersampling rates and graph densities, comparing the original approach with our improved sRASL-based method. 2 Background Causal inference from time series data, particularly from fMRI, has been a central focus in computational neuroscience due to its potential to uncover complex neural mechanisms. A directed dynamic causal model extends standard causal models ( Pearl and others 2000 ; Spirtes et al. 1993 ) by incorporating temporal dependencies, where the graph G includes nodes representing random variables \textbf{V}=\{V_{1},V_{2},\dots,V_{n}\} across both the current timestep t ( \textbf{V}^{t} ) and previous timesteps ( \textbf{V}^{t-k} ) that are direct causes of some V_{i}^{t} . These models assume a first-order Markov structure, where \textbf{V}^{t} is conditionally independent of \textbf{V}^{t-k} given \textbf{V}^{t-1} for all k>1 , allowing flexibility in the causal timescale ( Spirtes et al. 2000 ) . Undersampling presents a significant challenge in fMRI, where data is collected at intervals longer than the timescale of underlying neural processes. If the true causal structure evolves over timepoints \{t^{0},t^{1},t^{2},\dots,t^{k},\dots\} , but measurements are taken at \{t^{0},t^{u},t^{2u},\dots,t^{ku},\dots\} , this results in an undersampled graph \textbf{G}^{u} that may obscure the true causal graph \textbf{G}^{1} ( Danks and Plis 2013 ; Gong et al. 2015 ) . To address the structural implications of undersampling, several approaches have been developed. For instance, Danks and Plis 2013 explored the problem structurally, while Gong et al. 2015 provided a parametric approach for two-variable systems. The resulting undersampled graphs are typically compressed, with temporal information encoded in the edges. Specifically: (1) In the undersampled graph \mathcal{G}^{u} , an edge V_{i}\rightarrow V_{j} exists if and only if there is a directed path of length u from V_{i} to V_{j} in the fully observed graph \mathcal{G}^{1} , or equivalently, if there exists a directed path from V_{i}^{t-u} to V_{j}^{t} in G 1 .

Read the complete text