Limitations of Dynamic Causal Modelling for Multistable Cortical Circuits

Dynamic causal modelling (DCM) has been used extensively for inferring effective connectivity from neuroimaging data. However, it is unclear whether the DCM approach can reveal neural circuits with multistable dynamic states with global dynamical structures. In this work, we used excitatory-inhibitory cortical columnar neural mass models endowed with multistable dynamical states as ground truths to evaluate whether DCM can accurately identify their models’ architecture and connectivity strengths. Specifically, we simulated three recurrently connected neural mass models with different types of multistability, and generated local field potential data for DCM. The first model has bistable fixed points and exhibiting binary decision-making behaviour, and the second model has co-existence of two stable oscillatory frequencies (period doubling). The third model exhibits deterministic chaos, with a continuum of confined states. For each of the three models, DCM’s Bayesian model selection was able to correctly identify the correct model architecture among various options in model space. However, DCM’s Bayesian model averaging of the winning model was unable to accurately elucidate all the models’ extrinsic connectivity strengths, leading to the behaviours of the reconstructed models to differ substantially from their corresponding ground-truth models. Further, we found DCM’s estimation was highly sensitive to the sampling frequency used during its training. Overall, this work reveals the limitations of DCM in evaluating complex, multistable neural dynamical states and hence caution its use under these conditions.