Limitations of Variational Laplace-based Dynamic Causal Modelling for Multistable Cortical Circuits

Abstract

Dynamic causal modelling (DCM) is widely used to infer effective connectivity from neuroimaging data. However, its applicability to neural systems with complex, multistable dynamics remains uncertain--particularly when using the standard estimation approach based on variational Bayesian inference under the Laplace approximation. To investigate this limitation, we constructed biologically grounded cortical columnar neural mass models exhibiting three distinct multistable regimes: bistable fixed points associated with decision-making, coexisting oscillatory states through period-doubling bifurcations, and deterministic chaotic dynamics. These models were used to simulate local field potentials, which served as inputs to DCM. Bayesian model selection successfully identified the correct model architecture in all cases. However, Bayesian model averaging of the winning models failed to accurately estimate extrinsic connectivity parameters, leading to substantial discrepancies between the dynamics of the reconstructed and ground-truth systems. These results suggest that even when model selection is accurate, parameter estimation can break down under complex dynamics. Compared to previous applications of DCM to simpler neural systems, our study highlights significant limitations in its ability to capture the structure of multistable and globally nonlinear dynamics. We conclude that caution is warranted when applying variational Laplace-based DCM procedures to experimental paradigms involving bifurcations, chaotic trajectories, or other forms of dynamical complexity.