Can dynamic causal modelling (DCM) identify multistable neural circuits for decision-making?

Dynamic Causal Modeling (DCM) is commonly used for identifying neural circuit models from neuroimaging data to deduce causal neural interactions (Penny et al., 2011). Although previous studies have applied DCM to decision-making tasks, it is unclear whether DCM can reveal multistable dynamic states associated with choices (Wong and Wang, 2006). Here, we developed a two-column cortical fully self-feedback neural mass model (FSM) comprising two excitatory and one inhibitory neural populations, and inputs to both inhibitory and excitatory populations (Roach et al., 2023) with lateral connections between the columns. This network was then simulated to perform two-alternative decision-making and decision storage via multistable dynamics. Defining a decision threshold (columnar output difference), we categorised the outputs into correct, error and non-decision trials. Using the simulated model as ground truth, we estimated two types of DCMs, one with forward-backward connections and another with lateral connections between the columns using the SPM toolbox (local field potential). Bayesian model selection indicated that the expected and exceedance probability of the forward-backward model was higher for all the trial types and DCM approach struggled to identify the winning model. This suggests that while DCM is effective in many scenarios, it might require modifications to improve its predictive accuracy under multistable dynamic cognitive states. Future work will focus on evaluating different output measures and the re-evaluation of DCM’s parameterisation procedure.