Latent brain state dynamics distinguish behavioral variability, impaired decision-making, and inattention

Attention Deficit Hyperactivity Disorder (ADHD) affects 5–10% of children and is marked by heterogeneity in symptoms and developmental trajectories. A robust behavioral phenotype related to attentional deficits is intra-individual response variability (IIRV), reflecting trial-to-trial variability in response times. Prior work implicates large-scale brain networks—the salience network (SN), frontoparietal network (FPN), and default mode network (DMN)—in attention and cognitive control, with abnormalities reported in ADHD. Models suggest that fluctuations in latent brain states may underlie unstable performance and attentional lapses, but have not been rigorously tested with computational approaches. This study applies a Research Domain Criteria framework, using ultrafast task-fMRI and a Bayesian switching linear dynamic systems (BSDS) model to identify latent brain states and their temporal dynamics during a simple choice response task in children spanning ADHD and typically developing (TD) groups. Behavioral variability is quantified via ex-Gaussian parameters (std, tau), and latent decision processes are modeled with hierarchical drift-diffusion modeling (HDDM), focusing on whether engagement of task-optimal brain states relates to reduced IIRV, improved evidence accumulation (drift rate), and lower inattention, and whether non-optimal states relate to impaired decision-making and elevated inattention.

Participants: Fifty-two children participated (29 ADHD: 11 female/18 male; 23 TD: 11 female/12 male). Groups did not differ in age, gender, or head motion. ADHD group had higher inattention and hyperactivity/impulsivity scores. Task and behavioral measures: A simple choice response task was performed during fMRI. Children pressed left/right buttons to left-/right-pointing arrows with jittered inter-trial intervals. Accuracy and reaction time (RT) were recorded. IIRV was quantified using both Gaussian (RT standard deviation) and ex-Gaussian parameters (sigma, tau) to capture skewed RT distributions. Drift-diffusion modeling (HDDM): A hierarchical drift-diffusion model estimated latent decision parameters per child: decision boundary (a), drift rate (v), and non-decision time (t). Hypotheses focused on attention impacting drift rate (evidence accumulation speed) and impulsivity influencing decision boundary. fMRI acquisition and ROIs: Ultrafast task-fMRI with temporal resolution of 490 ms was collected. Regions of interest encompassed key nodes of SN (anterior insula, dorsal medial prefrontal/anterior cingulate), FPN (middle frontal gyrus, frontal eye fields, inferior parietal lobule), and DMN (posterior cingulate cortex, ventromedial prefrontal cortex), defined from an independent study showing attentional load effects. Latent brain state modeling (BSDS): Time series from ROIs were analyzed with a Bayesian switching linear dynamic systems model. BSDS performs unsupervised learning to infer hidden brain states and transitions without predefined temporal windows, modeling time-varying activation/connectivity in a latent subspace and using a hidden Markov model on latent variables. BSDS identified four latent brain states (S1–S4). Temporal properties included occupancy rate (percentage of time in a state) and mean lifetime (average dwell time before switching). Analytic strategy: Dimensional analyses related behavior (accuracy, RT metrics, HDDM parameters) and symptoms (inattention, hyperactivity/impulsivity) to state dynamics (occupancy, dwell time). Categorical analyses compared ADHD vs TD on behavior and state metrics. Associations between network connectivity (SN–FPN, SN–DMN) and behavioral/clinical indices were evaluated, focusing on links to drift rate and inattention.

- Behavioral performance (dimensional): Accuracy negatively correlated with inattention (r = -0.31, p = 0.02); not with hyperactivity/impulsivity. Mean RT was not significantly related to either symptom dimension. - Behavioral performance (categorical): ADHD showed lower task accuracy than TD (t = 3.08, p = 0.004); no group difference in average RT. - IIRV (dimensional): RT standard deviation (std) marginally correlated with inattention (r = 0.27, p = 0.05); RT tau marginally correlated with inattention (r = 0.25, p = 0.06); sigma not related to symptoms. - IIRV (categorical): ADHD > TD in RT std (t = 2.57, p = 0.01) and RT tau (t = 2.00, p = 0.05); no group difference in sigma. - Drift-diffusion modeling: Drift rate (v) negatively correlated with RT tau (r = -0.34, p = 0.01). Drift rate (v) negatively correlated with inattention (r = -0.28, p = 0.04), but decision boundary (a) and non-decision time (t) were not correlated with symptoms. Group differences showed ADHD < TD in drift rate (t = 2.95, p = 0.005) and ADHD > TD in non-decision time (t = 2.7, p = 0.01); no group difference in decision boundary (p = 0.2). - Latent brain states: BSDS identified four states (S1–S4). Occupancy rate of S1 was higher in TD than ADHD (TD: 21 ± 9% vs ADHD: 15 ± 8%; p < 0.05 FDR). Occupancy of S2 was higher in ADHD than TD (ADHD: 36 ± 13% vs TD: 27 ± 10%; p < 0.05 FDR). No group differences in mean lifetimes. - Brain states and IIRV: S1 occupancy was negatively correlated with RT std (r = -0.35, p = 0.008) and RT tau (r = -0.33, p = 0.02). S2 occupancy was positively correlated with RT std (r = 0.27, p = 0.05) and RT tau (r = 0.27, p = 0.046). Mean lifetimes of S1 and S2 were not related to RT variability. - Brain states and symptoms: S2 occupancy positively correlated with inattention (r = 0.27, p = 0.04), though not significant after multiple-comparisons correction; no relations with hyperactivity/impulsivity. - Network connectivity-behavior links (from abstract): SN–FPN functional connectivity predicted drift rate (evidence accumulation), whereas SN–DMN connectivity predicted inattention. - Overall: A task-optimal state (S1) was linked to reduced behavioral variability and enhanced evidence accumulation; a non-optimal state (S2) was linked to inattention and differentiated ADHD from TD in categorical analyses.

The study addressed whether latent brain state dynamics within canonical attention-control networks (SN, FPN, DMN) explain behavioral variability, decision-making efficiency, and inattention in children. Using BSDS on ultrafast task-fMRI, a task-optimal state (S1) was identified whose higher occupancy related to lower intra-individual response variability and better decision evidence accumulation, aligning with models positing SN/FPN engagement supports stable attention and efficient control. Conversely, a non-optimal state (S2) showed higher occupancy in ADHD, positive associations with IIRV, and links to inattention severity, supporting the view that altered brain state dynamics underlie attentional fluctuations in ADHD. Drift-diffusion modeling demonstrated that inattention specifically relates to reduced drift rate, indicating compromised evidence accumulation, and this cognitive deficit was further tied to large-scale network interactions (SN–FPN predicting drift rate; SN–DMN predicting inattention). Together, these results provide convergent evidence that distinct latent brain state properties and network connectivity patterns dissociate behavioral instability, impaired decision-making, and inattentional symptoms, offering mechanistic insight into ADHD-related attentional dysfunction.

This work integrates computational psychiatry and systems neuroscience by combining BSDS-based latent state discovery with HDDM of decision processes in children performing a simple choice task. Findings reveal dissociable latent brain state dynamics that map onto response variability, evidence accumulation efficiency, and inattention, and distinguish ADHD from TD children at the level of state occupancy. The results advance understanding of the neurobiological underpinnings of attention deficits in ADHD and highlight the relevance of SN–FPN and SN–DMN interactions for decision-making and symptom severity.