Critical dynamics predicts cognitive performance and provides a common framework for heterogeneous mechanisms impacting cognition

Cognitive function emerges from cortical network structure and dynamics and is frequently impaired in neuropsychiatric disorders, notably epilepsy. Persons with epilepsy exhibit heterogeneous cognitive profiles with common impairments in language, memory, attention, and executive function, arising from multifactorial causes including disease etiology and dynamic influences such as ASMs, interictal/ictal activity, and disrupted sleep. Physics and information theory propose that cortical networks operate optimally at a critical state—at the phase transition between ceasing and chaotic activity—where excitation and inhibition are balanced and information transmission, integration, storage, dynamic range, and sensitivity are simultaneously optimized. Hallmarks of critical dynamics include power-law scaling and long-range temporal correlations (TCs). Prior work suggests that criticality signatures are perturbed by ASMs, sleep/wake pressure, and deprivation. The central research question here is whether proximity to critical dynamics predicts global cognitive performance in humans and whether heterogeneous mechanisms (IEDs, ASMs, SWSs) affect cognition by perturbing criticality.

Evidence for brain criticality includes neuronal avalanches and power-law scaling in animal and human recordings, and long-range temporal correlations reflecting intrinsic timescales across cortex. Studies have associated critical-like dynamics with specific cognitive processes such as motor learning, attention, executive function, and cognitive flexibility, and proposed their utility as biomarkers in ADHD and Alzheimer’s disease. However, most prior work relied on noninvasive methods (EEG, MEG, fMRI) with limited high-γ sensitivity, short recordings, coarse spatial coverage, and susceptibility to artifacts, and often focused on single tasks rather than comprehensive cognitive profiling. Criticality signatures are known to change systematically with pharmacological perturbations (e.g., ASMs) and vigilance/sleep pressure, supporting a mechanistic link between network state and information processing, yet the relationship to broad cognitive function and the convergence of heterogeneous clinical factors through criticality remained unclear.

Participants: Retrospective analysis of drug-resistant focal epilepsy patients undergoing presurgical intracranial video-EEG monitoring. Two cohorts: Dataset 1—81 PwE (35 females, mean age 32.3 ± 10.7 years) from Berlin-Brandenburg Epilepsy-Center; Dataset 2—23 PwE (12 females, mean age 28.5 ± 13.2 years) from University of Freiburg Epilepsiae database. Clinical characteristics included epilepsy duration, seizure history, MRI findings, epilepsy localization, seizure onset side, and surgical outcomes.

Cognitive Testing: Performed in all 81 subjects (dataset 1), covering four domains: language, attention, working/short-term memory, and verbal learning/memory via 14 tests during routine neuropsychological evaluation (mean 6.8 ± 4.1 months prior to iEEG). Domain impairment defined if two parameters within a domain scored ≥1 SD below population norms.

ASM Quantification: Daily ASMs normalized by defined daily dose (DDD) and summed per day. High vs low ASM days identified per patient undergoing tapering; days with rescue benzodiazepines excluded. Subjects without tapering excluded from ASM analyses.

iEEG Preprocessing: Dataset 1—5 min per hour across multiday monitoring; Dataset 2—continuous data from two days. Preprocessing: notch filters at 50/100 Hz; downsampled to 256 Hz with antialiasing; bandpass 0.1–128 Hz; artifact channels removed via criteria (constant signal or abnormal frequency peaks >6× IQR); visual inspection; seizure segments excluded with ±10 min pre/post margins per clinical marks.

Temporal Correlations (TCs): High-γ power (56–96 Hz) estimated via Welch’s method (Hann window) in nonoverlapping 125 ms windows; powers log10-normalized. Autocorrelation functions computed in consecutive 120 s windows (90 s overlap) per channel; TCs defined as the lag where autocorrelation first falls below half the difference between first lag and baseline (median 40–60 s). Lag range ~0.125–60 s; lag zero excluded. Autocorrelation functions aggregated by median over windows. For global state comparisons (e.g., ASM days), subsampling to minority group size, repeated 100× and averaged. Surrogate TCs derived from time-shuffled high-γ power preserving power distributions but destroying temporal relations.

SWS Detection: SWS scored in 30 s windows using a validated algorithm based on a vigilance index ((θ+δ)/(α+β_high+spindle)). SWS defined as >1 SD above daily mean. 120 s windows marked SWS if any of the four 30 s subwindows flagged; otherwise nonSWS.

IED Detection: Automated deep-learning pipeline: template matching to identify candidate IEDs, spectrogram transformation for dimensionality reduction, pretrained DNN classification. TCs compared in segments/channels without IEDs and with 5–30 IEDs per minute; channels with <50 segments in both bins excluded.

Spectral Power: Median power per 120 s segment via Welch’s method in δ (0.5–4 Hz), θ (4–8 Hz), α (8–12 Hz), β (12–30 Hz), γ (30–45 Hz), and high-γ (55–95 Hz) bands.

Statistical Analysis: Paired Wilcoxon tests for within-patient state comparisons (ASM, SWS, IED bins). Brunner–Munzel tests for between-patient comparisons relating EEG features to cognition given nonnormal, heteroscedastic TC distributions. Multiple comparisons controlled via Benjamini–Hochberg (α=0.05). Effect sizes reported as nonparametric relative effects. P-value distributions tested against uniform via Kolmogorov–Smirnov.

Neural Network Modeling: Discrete-time network of N=1024 neurons on a 2D grid with periodic boundaries, 20% inhibitory, 80% excitatory, all-to-all connectivity with distance-dependent Gaussian scaling exp(−r^2/(2σ^2)), σ=4; self-connections omitted. Weight matrix w scaled so that largest absolute eigenvalue λ serves as control parameter for distance to criticality. Neuron i active s_i(t)=1 or inactive s_i(t)=0; activation probability at t+1 determined by summed input with thresholds (0 if sum<0; 1 if sum>1; else sum). Background activity enforced by randomly activating one neuron at each timestep for max_t=5000 timesteps; simulations repeated 1000×; TCs extracted from autocorrelation of average firing after 500 timesteps. Perturbations: ASMs modeled by reducing outgoing excitatory strengths by factor f_exc; SWS by off-periods where all neurons remain quiescent with probability P_off; IEDs by probabilistic local spikes activating a random neuron and neighbors (20% of neurons) with probability P_IED.

Ethics: Approved by Charité IRB (dataset 1; consent waived due to retrospective design); University of Freiburg IRB and written consent for Epilepsiae database (dataset 2).

Model results: TCs peak near the critical point (λ≈1) and are short in subcritical (λ<1) and supercritical (λ>1) regimes. Increasing SWS off-period probability (P_off), IED probability (P_IED), or reducing excitatory strength (f_exc, modeling ASMs) shortens TCs, with strongest effects near the critical regime.

Human iEEG (104 PwE) findings:

• SWS prevalence: Dataset 1 SWS epochs constituted 18±7% of segments; Dataset 2 19±6%.
• SWS effect on TCs (excluding IED segments): Dataset 1—nonSWS TC = 0.53±0.13 s vs SWS TC = 0.45±0.08 s, P<0.001; Dataset 2—nonSWS TC = 1.0±0.5 s vs SWS TC = 0.8±0.5 s, P<0.01.
• IED rates: Dataset 1 average 2.6±2.4 IEDs/(channel·min); Dataset 2 3.1±1.7 IEDs/(channel·min).
• IED effect on TCs (excluding SWS segments): Dataset 1—no IED TC = 0.6±0.4 s vs 5–30 IEDs/min TC = 0.5±0.2 s, P<0.001; Dataset 2—no IED TC = 1.0±0.5 s vs 5–30 IEDs/min TC = 0.8±0.3 s, P<0.05.
• ASM tapering effects (excluding SWS and IED segments): Dataset 1 (N=60; 50±30% reduction)—low ASM TC = 0.53±0.12 s vs high ASM TC = 0.48±0.11 s, P<0.01; Dataset 2 (N=23; 70±30% reduction)—low ASM TC = 1.2±0.7 s vs high ASM TC = 0.9±0.4 s, P<0.01.
• SOZ vs non-SOZ: SWS, IEDs, and ASMs shorten TCs in both SOZ and non-SOZ, indicating widespread network effects; trends suggest longer TCs in SOZ when accounting for IEDs (significant in dataset 1 only).
• Cognition: Across 102 subcomparisons relating EEG/clinical features to cognitive domains, 20 showed P<0.05 and 11 remained significant after BH correction, with TCs the only robust predictor across attention, language, and working memory domains. IED rates, ASM load, SWS proportion, and spectral powers did not predict cognition after correction. TC P-value distribution significantly deviated from uniform (KS P<0.01), indicating non-null associations; strongest correlations observed on first recording day.
• Controls: Surrogate TCs showed no such effects; raw high-γ power and SNR did not explain TC changes. Hurst exponents and α-band TCs showed similar trends but were less robust than high-γ TCs.

The findings support the brain criticality hypothesis by demonstrating that proximity to critical dynamics, measured via temporal correlations of high-γ activity, predicts cognitive performance across multiple domains. Heterogeneous mechanisms known to influence cognition—SWS, IEDs, and ASM load—converge on a common endpoint: perturbation away from the critical state and shortening of TCs. This unifying framework explains how diverse clinical factors impair information processing, integration, and storage by altering network dynamics. The strongest cognitive associations on day one likely reflect minimal confounding from hospitalization-related changes (medication adjustments, accumulating seizures, variable monitoring duration). Analyses across SOZ and non-SOZ suggest that these dynamics are widespread and not limited to epileptogenic regions; trends of longer TCs in SOZ (after accounting for IEDs) may indicate closer proximity to criticality and vulnerability to supercritical transitions associated with seizure initiation. Importantly, spectral power measures, IED rates, ASM dosage, and SWS proportion alone did not predict cognition after correction, underscoring the specificity of TC-based criticality metrics. These results highlight criticality-informed measures as potentially valuable biomarkers for cognitive state, with broader implications beyond epilepsy, and emphasize the need to balance optimization of network function with seizure susceptibility near criticality.

This study demonstrates that temporal correlations, a hallmark of critical dynamics, robustly predict cognitive performance profiles in persons with epilepsy. Neural network modeling and multiday intracranial EEG reveal that SWS, interictal epileptiform discharges, and antiseizure medications consistently shorten TCs by shifting dynamics away from the critical state. Critical dynamics thus provide a common framework and setpoint for optimal cortical network function and cognition, integrating heterogeneous mechanisms impacting cognition. Future work should investigate prospective, concurrent cognitive assessments with EEG, extend analyses to other neurological and psychiatric conditions, develop interventions to modulate network criticality safely, and explore TC-based biomarkers for monitoring cognitive health while carefully managing seizure risk inherent to operating near criticality.

• Recording design: Dataset 1 comprised 5-minute iEEG segments each hour rather than continuous monitoring, though compensated by long monitoring durations and large sample size.
• Automated detections: SWS and IEDs were detected using validated algorithms without manual annotations; full polysomnographic sleep staging was not possible due to lack of EMG/scalp EEG.
• Temporal mismatch: Cognitive testing occurred months prior to iEEG and not concurrently, potentially introducing bias given dynamic cognitive states and ASM changes, though mitigated by comprehensive testing.
• Spatial coverage: Intracranial electrodes predominantly sampled temporal and frontal regions, which may bias domain-specific findings (e.g., language, attention).
• Interpretation of IED-related criticality shifts: While increased IEDs shorten TCs, the precise direction of deviation (subcritical vs supercritical) cannot be conclusively determined from the present data and may vary across contexts.