Modeling Behavioral and Neuroimaging Dynamics

This project develops innovative methodologies and scalable computational frameworks for modeling human decision-making dynamics, learning robust and interpretable representations from neuroimaging data.

We proposed a novel framework that integrates reinforcement learning (RL), drift-diffusion models (DDMs), and hidden Markov models to jointly analyze choices, response times, and latent strategy switching. Furthermore, we introduced a multi-task learning framework to simultaneously model multiple behavioral tasks, enabling information to be shared across tasks and allowing each task to benefit from insights gained from the others. Our framework revealed that patients with major depressive disorder (MDD) show lower overall engagement and reduced focus than healthy controls, and take longer to make decisions when engaged and focused. Additionally, we find that for MDD patients, neuroimaging measures of brain activity are associated with decision-making characteristics in the engaged state, but not in the lapsed state, providing evidence for brain–behavior associations specific to engagement. Whlie the observed response times do not predict treatment response, shared parameters identified by our framework are predictive of treatment responses, demonstrating potential as new behavioral markers of therapeutic outcomes.

Standard RL relies on prespecified update rules, whereas vanilla neural networks lack subject- and state-specific learning mechanisms. We propose a personalized state-partitioned gated recurrent unit (GRU) model, which outperforms asymmetric Q-learning and RL-DDM. In addition, the GRU hidden states recover Q-value–like signals without requiring explicit update rules, and the inferred subject-level deviations align with known MDD biomarkers.

Currently, we are developing methods to jointly model task-neuroimaging data to uncover latent neural dynamics and elucidate brain–behavior relationships. We are also exploring neuroimaging foundation models to facilitate subgrouping in MDD.