Rubén Herzog

A main characteristic of biological systems is their capacity to dynamically adapt to environmental changes. In the brain, synaptic plasticity enables the strengthening or weakening of connections between neurons, allowing neural circuits to adapt based on experience, learning, and environmental changes. Yet, it is homeostatically regulated such that it avoids excessive proliferation of synaptic contacts. These mechanisms can be studied with large-scale models of brain activity. Here, we embed a biologically grounded inhibitory-homeostatic plasticity rule into the Dynamic Mean Field (DMF) model, creating a Homeostatic Dynamic Mean Field (HDMF) model that dynamically tunes local excitation–inhibition balance. Convergence of excitatory firing rates is reached by mapping a large range of coupling strength to parameters of inhibitory synapses. The HDMF reproduces statistical observables of brain activity as well as the original DMF, and can sustain neuromodulatory perturbations without overhead computations. The HDMF can generate unprecedented sleep-like slow-wave activity, which can also coexist with wake-like asynchronous dynamics, permitting to model dissociated states of consciousness such as parasomnias. Together, these results show that a single homeostatic rule broadens the stability and expressiveness of the DMF, providing a unified platform for studying how local adaptive processes shape the diverse global dynamics of the human brain.

Ketamine is a fast-acting intervention for treatment-resistant depression (TRD), yet only a subset of patients show robust clinical response, and the underlying neural mechanisms remain unclear. High-order interactions (HOI) derived from multivariate information theory provide a framework for examining nonlinear dependencies among brain regions beyond pairwise connectivity. One such metric, the O-information, captures the balance between synergistic and redundant interactions across three or more variables. In this secondary analysis of a randomized, double-blind, midazolam-controlled trial (NCT02556606), we examined EEG-derived HOI in 30 late-life veterans with TRD following a single 40-minute intravenous infusion of ketamine (0.1, 0.25, 0.5 mg/kg; n = 18) or midazolam (0.03 mg/kg; n = 12). Resting state and mismatch negativity data were analyzed at baseline, 1 h, 24 h, and 7 d post-infusion. Ketamine induced temporally dynamic alterations in redundancy-dominant O-info, with maximal effects in the alpha-band at 1 h (Cohen’s d = 2.57), attenuation at 24 h that shifted toward the theta-band, and partial resurgence in beta and gamma by Day 7. Linear mixed-effects modeling identified significant group effects across most band x metric families, with the strongest effects in alpha, beta, and gamma redundancy. Greater increases in 24-hour alpha-band redundancy were associated with greater improvement in depressive symptoms at Day 7 (β = 69.31, q = 0.05). HOI metrics also tracked acute dissociative states, with several 24-hour alpha and beta features remaining positively associated with symptom severity after correction. These findings extend prior HOI work in healthy samples to a controlled TRD cohort and suggest that ketamine induces temporally structured reorganization of higher-order brain interactions, with exploratory associations to clinical outcomes.

INTRODUCTION: Growing evidence supports a critical role of sleep slow waves (SW) in Alzheimer's disease (AD). However, wake SW (sleep‐like SW potentially reflecting local intrusions of sleep) remain unexplored in AD. METHODS: A total of 274 older adults with subjective cognitive decline (SCD) (INSIGHT‐preAD cohort, 76.6 ± 3.5 years) underwent (i) positron emission tomography (PET) scans for amyloid (A) and neurodegeneration (N), (ii) high‐density resting‐state electroencephalogram (EEG) recordings to detect wake SW, and (iii) cognitive assessments. PET biomarkers were reassessed 2 years later. We examined wake SW associations with (1) current A/N status, (2) cognition, and (3) amyloid conversion. RESULTS: A+N−, A−N+, and A+N+ individuals exhibited lower delta wake SW density than A−N− participants. Wake SW amplitude (1) was higher in A+N+ than A−N− individuals, (2) correlated with poorer memory, and (3) predicted A− to A+ conversion ( n = 157 A‐ stable individuals, n = 16 convertors). DISCUSSION: Wake SW represent promising early EEG biomarkers for AD pathology and amyloid conversion, facilitating risk stratification before cognitive decline onset.

Complex systems are characterized by nonlinear dynamics, multi-level interactions, and emergent collective behaviors. Traditional analyses that focus solely on pairwise interactions often oversimplify these systems, neglecting the higher-order interactions critical for understanding their full collective dynamics. Recent advances in multivariate information theory provide a principled framework for quantifying these higher-order interactions, capturing key properties such as redundancy, synergy, shared randomness, and collective constraints. However, two major challenges persist: accurately estimating joint entropies and addressing the combinatorial explosion of interacting terms. To overcome these challenges, we introduce THOI (Torch-based High-Order Interactions), a novel, accessible, and efficient Python library for computing high-order interactions in continuous-valued systems. THOI leverages the well-established Gaussian copula method for joint entropy estimation, combined with state-of-the-art batch and parallel processing techniques to optimize performance across CPU, GPU, and TPU environments. Our results demonstrate that THOI significantly outperforms existing tools in terms of speed and scalability. Specifically, THOI reduces the time required to exhaustively analyze all interactions in small systems (≤ 30 variables). For larger systems, where exhaustive analysis is computationally impractical, THOI integrates optimization strategies that make higher-order interaction analysis feasible. We validate THOI’s accuracy using synthetic datasets with parametrically controlled interactions and further illustrate its utility by analyzing fMRI data from human subjects in wakeful resting states and under deep anesthesia. Finally, we analyzed over 900 real-world and synthetic datasets, establishing a comprehensive framework for applying higher-order interaction (HOI) analysis in complex systems. THOI opens new perspectives for testing both established and novel hypotheses about the multi-level, nonlinear, and multidimensional nature of complex systems.

INTRODUCTION: Growing evidence supports a critical role of sleep slow waves (SW) in Alzheimer's disease (AD). However, wake SW (sleep-like SW potentially reflecting local intrusions of sleep) remain unexplored in AD. METHODS: 274 older adults with subjective cognitive decline (INSIGHT-preAD cohort, 76.6 ± 3.5 years) underwent i) PET scans for amyloid (A) and neurodegeneration (N), ii) high-density resting-state EEG recordings to detect wake SW, and iii) cognitive assessments. Biomarkers were reassessed two years later. We examined wake SW associations with 1) current A/N status, 2) cognition, and 3) amyloid conversion. RESULTS: A+N−, A−N+ and A+N+ individuals exhibited lower delta wake SW density than A−N− participants. Wake SW amplitude 1) was higher in A+N+ than A−N− individuals, 2) correlated with poorer cognition, and 3) predicted A− to A+ conversion. DISCUSSION: Wake SW represent promising early EEG biomarkers for AD pathology and amyloid conversion, facilitating risk stratification before cognitive decline onset.

Magneto/electroencephalography (M/EEG) studies of dreaming are an essential paradigm in the investigation of neurocognitive processes of human consciousness during sleep, but they are limited by the number of observations that can be collected per study. Dream research also involves substantial methodological and conceptual variability, which poses problems for the integration of results. To address these issues, here we present the DREAM database—an expanding collection of standardized datasets on human sleep M/EEG combined with dream report data—with an initial release comprising 20 datasets, 505 participants, and 2643 awakenings. Each awakening consists, at minimum, of sleep M/EEG ( ≥ 20 s, ≥100 Hz, ≥2 electrodes) up to the time of waking and a standardized dream report classification of the subject’s experience during sleep. We observed that reports of conscious experiences can be predicted with objective features extracted from EEG recordings in both Rapid Eye Movement (REM) and non-REM (NREM) sleep. We also provide several examples of analyses, showcasing the database’s high potential in paving the way for new research questions at a scale beyond the capacity of any single research group.

Neuroimaging techniques produce vast amounts of data, capturing brain activity in a high-dimensional space. However, brain dynamics are consistently shown to reside in a rather lower-dimensional space, which contains relevant information for cognition and behavior. This dimensionality reduction reflects distinct types of interactions between brain regions, such as redundancy –shared neural information distributed across regions– and synergy, where information emerges only when regions are considered collectively. Significant efforts have been devoted to developing linear and non-linear algorithms to reveal these low-dimensional dynamics, often termed "brain modes." Here, we apply various dimensionality reduction techniques to resting-state functional magnetic resonance imaging (fMRI) data from 100 healthy participants to examine how synergistic interactions in brain dynamics are preserved by these techniques. We first demonstrate that biologically informed brain parcellation modulates and preserves synergy-dominated interactions. Next, we show that synergy among low-dimensional modes enhances functional-connectivity reconstruction: nonlinear autoencoders not only achieve the lowest reconstruction error but also maximally preserve synergy, outperforming principal component analysis, diffusion maps, and Laplacian eigenmodes. Finally, we confirm previous results suggesting that global signal regression helps to identify synergistic interactions between regions. Our findings establish synergy preservation as a complementary criterion to reconstruction accuracy, highlighting autoencoders as a nonlinear tool for uncovering synergistic low-dimensional brain modes from high-dimensional neuroimaging data.

Neuroimaging techniques produce vast amounts of data, capturing brain activity in a high-dimensional space. However, brain dynamics are consistently shown to reside in a rather lower-dimensional space, which contains relevant information for cognition and behavior. This dimensionality reduction reflects distinct types of interactions between brain regions, such as redundancy -shared neural information distributed across regions- and synergy, where information emerges only when regions are considered collectively. Significant efforts have been devoted to developing linear and non-linear algorithms to reveal these low-dimensional dynamics, often termed "brain modes." Here, we apply various dimensionality reduction techniques to resting-state functional magnetic resonance imaging (fMRI) data from 100 healthy participants to examine how synergistic interactions in brain dynamics are preserved by these techniques. We first demonstrate that biologically informed brain parcellation modulates and preserves synergy-dominated interactions. Next, we show that synergy among low-dimensional modes enhances functional-connectivity reconstruction: nonlinear autoencoders not only achieve the lowest reconstruction error but also maximally preserve synergy, outperforming principal component analysis, diffusion maps, and Laplacian eigenmodes. Finally, we confirm previous results suggesting that global signal regression helps to identify synergistic interactions between regions. Our findings establish synergy preservation as a complementary criterion to reconstruction accuracy, highlighting autoencoders as a nonlinear tool for uncovering synergistic low-dimensional brain modes from high-dimensional neuroimaging data.

Understanding insomnia is crucial for improving its diagnosis and treatment. However, many subjective complaints about insomnia do not align with objective measures of sleep quality, as is the case in subjective-objective sleep discrepancy (SOSD). We address this discrepancy by measuring sleep intrusions and instability in polysomnographic recordings from a large clinical database. Using machine learning, we develop personalized models to infer hypnodensities—a continuous and probabilistic measure of sleep dynamics—, and analyze them via information theory to measure intrusions and instability in a principled way. We find that insomnia with SOSD involves sleep intrusions during intra-sleep wakefulness, while insomnia without SOSD shows wake intrusions during sleep, indicating distinct etiologies. By mapping these metrics to standard sleep features, we provide a continuous and interpretable framework for measuring sleep quality. This approach integrates and values subjective insomnia complaints with physiological data for a more accurate view of sleep quality and its disorders.