STUDY OF BRAIN DYNAMICS IN COMPLEX AND DYNAMIC EXPERIMENTS

Abstract

This dissertation advances analytical frameworks for studying brain dynamics in complex fMRI experiments. Modern neuroscience is shifting toward dynamic paradigms that better mimic real-world scenarios, but these pose fundamental analytical challenges. First, they often lack the well-defined events or trials that serve as basic units of analysis. Second, they challenge the core assumption of trial-averaging. Recent work suggests that trial-to-trial variability is not simply noise, but a critical signal corresponding to meaningful behaviors or internal states. Conventional methods, which discard this variability, are thus ill-suited for these experiments. This thesis addresses these challenges by developing and applying a suite of state-space and machine learning models to analyze brain activity as it unfolds dynamically. First, to establish a new unit of analysis, we apply Switching Linear Dynamical Systems (SLDS) to model fMRI data from a continuous threat-of-shock experiment. The model parsed fMRI timeseries into discrete brain states that successfully mapped onto experimental events (e.g., threat proximity). By separating the dynamics into intrinsic and extrinsic components, we show that these states evolve toward stable fixed-point attractors and quantify how external inputs steer the system’s trajectory Next, we extend this SLDS framework to test the hypothesis that trial-to-trial variability encodes rich internal processing of experimental stimuli. We represent this internal processing by characterizing each trial from a dynamic threat-avoidance task as a latent state sequence. Clustering these sequences reveals that a single experimental condition elicits a repertoire of distinct “dynamic modes,” each defined by a unique sequence of threat and safety states. These modes recurred across different experimental conditions, challenging the one-to-one stimulus-response assumption. We demonstrate these modesare functionally significant, proving predictive of physiological arousal (skin conductance) and corresponding to different ways of engaging with external stimuli, as revealed by a controllability analysis. Finally, this dissertation addresses the critical challenge of interpretability: linking the abstract, system-level findings of dynamical models (such as attractors and state trajectories) back to their concrete neurobiological substrates (such as specific brain regions). We first approach this in the SLDS framework by introducing a novel “region importance” measure to identify which brain areas are most influential in steering state trajectories . We then extend this focus on interpretability to a supervised recurrent neural network (RNN) framework, which we developed to model the spatiotemporal dynamics of highly complex naturalistic stimuli (movie-watching) for which unsupervised models are less optimal. This RNN approach uses saliency maps and lesion analyses to identify critical brain regions . We demonstrate these dynamic RNN representations are highly generalizable and rich enough to predict stable cognitive traits, such as fluid intelligence, from a participant’s brain response. Taken together, these studies provide a cohesive methodological progression for characterizing brain dynamics in complex experiments, moving beyond static analyses to capture meaningful, time-varying neural processes.