Experimental design and dynamic brain state modeling framework

(a) episodic memory task paradigm. During fMRI scanning, participants encoded 100 images of buildings and natural scenes (3 seconds each, jittered fixation 2-6 seconds), followed by a 6-minute post-encoding rest period. Participants were instructed to imagine themselves in each scene and memorize the images for subsequent testing. (B) Post-scan memory assessment. Outside the scanner, participants completed a memory retrieval test with confidence ratings using a 4-point scale (“Unsure” to “Very Sure”) to assess memory strength and subsequent memory effects. (C) Memory-related brain regions. Lateral view showing significant activation clusters (remembered > forgotten contrast) across ventral visual areas, medial temporal lobe, and frontal-parietal regions that served as the foundation for dynamic state modeling. (D) Regions of Interest for state space modeling. Sixteen core regions exhibiting significant subsequent memory effects, spanning visual association (VS) (fusiform, lateral occipital complex), memory encoding-related medial temporal lobe (MTL) (hippocampus, parahippocampal cortex), and cognitive control (frontal-parietal network) systems and default mode networks (DMN) (medial prefrontal and posterior cingulate cortex). ROIs are separately color coded by system. (E) Time series extraction. Representative raw BOLD time series from memory-related ROIs (Right anterior hippocampus, left anterior hippocampus, posterior cingulate cortex and medial prefrontal cortex) during encoding, preprocessed and normalized for dynamic brain state modeling. (F) Bayesian Switching Dynamic Systems model. Schematic illustration of the unsupervised BSDS framework that identifies latent brain states from multi-region time series without arbitrary parameter constraints. (G) State evolution visualization. Color-coded temporal evolution of four distinct brain states inferred by BSDS across all participants during memory encoding, revealing moment-to-moment state transitions at TR-level resolution.

Characterization of dynamic brain states during memory encoding.

(a) State-specific activation patterns. Polar plots depicting mean amplitude profiles across 16 memory-related regions for each of the four brain states identified by BSDS. State 3 (core-encoding state) shows prominent activation across visual areas, MTL system, and frontal-parietal regions. State 2 (default mode state) exhibits selective activation in medial prefrontal and posterior cingulate cortex. State 4 (inactive state) displays uniformly low activation, while State 1 shows intermediate patterns. (B) Functional connectivity matrices. Group-level covariance matrices revealing distinct connectivity signatures for each brain state. The active-encoding state (S3) demonstrates the highest overall functional connectivity, significantly exceeding all other states. (C-D) Network topology analysis. Bar graphs show clustering coefficient and global efficiency metrics derived from connectivity matrices. The active-encoding state exhibits significantly higher clustering coefficient and global efficiency compared to other states, indicating optimal balance between network integration and segregation for memory formation. Error bars represent standard error. Notes: *p < 0.05, **p < 0.01.

Brain state dynamics predict individual memory performance.

(A-B) State temporal properties. Box plots show occupancy rates and mean lifetimes for each brain state during encoding. The active-encoding state dominates with 38.39% occupancy, significantly higher than all other states. (C) Memory performance correlations. Scatter plots revealing that core-encoding state occupancy correlate positively with memory accuracy, while inactive state occupancy shows strong negative correlation. (D) State transitions. Network diagram illustrates transition probabilities between brain states during encoding, with edge thickness representing transition strength. Prominent bidirectional transitions occur between core-encoding, high-order control, and inactive states. (E-F) Transition-performance relationships. Scatter plots demonstrating that transitions toward the active-encoding state (S2→S3; S4→S3) enhance memory performance, while transitions toward the inactive state (S2→S4; S3→S4) impair performance. Self-transitions of the inactive state also negatively predict memory. Dashed lines indicate 95% confidence intervals; solid lines show best linear fit. Notes: ∼p < 0.10,*p < 0.05, **p < 0.01, ***p < 0.001.

Encoding state reinstatement during post-encoding rest supports memory consolidation.

(A) State decoding framework. Schematic workflow for detecting encoding state re-occurrence during post-encoding rest using the BSDS model trained on encoding data. Maximum log-likelihood estimation determines state assignment at each time point during rest. (B) State occupancy shifts. Bar graph comparing state occupancy between encoding and rest phases. All task-related states (S2, S3, S4) decrease significantly during rest, while State 1 increases substantially, reflecting the state dynamic shift during rest (C)Mean lifetime shifts. Bar graph comparing mean lifetime between encoding and post -encoding rest phases. All states decrease significantly during rest. (D) Memory-predictive reinstatement. Scatter plot showing that the mean lifetime of the high-order control state (S2) during post-encoding rest, but not pre-encoding rest, correlates strongly with memory accuracy, indicating that sustained activation of this MPFC-PCC network supports offline consolidation. (E) Rest transition dynamics. Network diagram showing altered transition patterns during rest, with decreased self-transitions and increased inter-state transitions compared to encoding. (F) Consolidation state stability. Scatter plot demonstrates that self-transitions of the default-mode state during post-encoding rest (not pre-encoding rest) show the strongest correlation with memory performance, suggesting that stability of this consolidation-related network configuration is critical for memory outcomes. Dashed lines indicate 95% confidence intervals; solid lines show best linear fit. Error bars represent standard error. Notes: *p < 0.05, ***p < 0.001.