A Data-Driven Closed-Loop Control Approach to Drive Neural State Transitions for Mechanistic Insight

Abstract

Repetitive negative thinking (RNT) is a transdiagnostic risk factor for mood disorders, consistently associated with altered biological substrates, including functional connectivity in key brain networks. As a stable cognitive feature linked to vulnerability across disorders, RNT presents a compelling target for intervention. However, leveraging RNT as a modifiable mechanism requires a deeper understanding of its causal neural dynamics and how targeted modulation can induce adaptive change. We introduce a data-driven framework that combines dynamical system reconstruction (DSR) with model predictive control (MPC) to infer optimal control policies for transitions between resting and sad mood brain states from functional magnetic resonance imaging (fMRI) data. Using nonlinear generative DSR models trained on individuals with remitted major depressive disorder (rMDD) and matched healthy controls (HCs), we derive region-specific, state-dependent control strategies. We find that small brain regions (e.g., sgACC, NAcc) exhibit higher controllability, requiring less energy to drive state transitions. Critically, rMDD participants require less control energy than HCs to move into sad mood from rest and - unexpectedly - also to move back to rest, though the latter effect is spatially restricted. Despite comparable target attainment, rMDD participants remain closer to the sad mood distribution when returning to rest, indicating a residual negative-affect bias. Our data-driven analysis reveals elevated effective coupling in rMDD, most prominently toward (but not away from) the DLPFC. Across regions, greater coupling is associated with reduced control energy, suggesting that enhanced network influence facilitates more efficient state transitions. These results suggest dynamics in rMDD that facilitate entry into negative affect and hinders full disengagement without sustained input, highlighting closed-loop control as a tool for mechanistic insight and potentially for designing targeted neuromodulatory interventions in the future.