Low-dimensional representation of brain networks for seizure risk forecasting

Steven Rico-Aparicio, Martin Guillemaud, Alice Longhena, Vincent Navarro, Louis Cousyn, Mario Chavez · 2025

Read Low-dimensional representation of brain networks for seizure risk forecasting by Steven Rico-Aparicio, Martin Guillemaud, Alice Longhena, Vincent Navarro, Louis Cousyn, Mario Chavez (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-00856.txt.

Introduction

Steven Rico-Aparicio and co-authors compress brain networks to forecast epilepsy seizures in a 2025 paper. The hours before a seizure, the preictal state, are the window where warnings could help, but the connectivity patterns that mark it are subtle and sprawled across many electrodes.

The team embeds functional brain networks, built from implanted electrode recordings, into a small Euclidean space that keeps their essential shape. A simple marker computed from the embedding separates seizure-free periods from the day before a seizure. A permutation test picks the most informative electrodes. Cross-validation and a mock forecasting run test whether the signal holds up across patients.

Catalog record

CREATORSteven Rico-Aparicio, Martin Guillemaud, Alice Longhena, Vincent Navarro, Louis Cousyn, Mario Chavez
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE45 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-00856

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Low-dimensional representation of brain networks for seizure risk forecasting Steven Rico-Aparicio, Martin Guillemaud, Alice Longhena, Vincent Navarro, Louis Cousyn, Mario Chavez Abstract Identifying preictal states—periods during which seizures are more likely to occur—remains a central challenge in clinical computational neuroscience. In this study, we introduce a novel framework that embeds functional brain connectivity networks, derived from intracranial EEG (iEEG) recordings, into a low-dimensional Euclidean space. This compact representation captures essential topological features of brain dynamics and facilitates the detection of subtle connectivity changes preceding seizures. Using standard machine learning techniques, we define a dimensionless biomarker, ℬ \mathcal{B} , that discriminates between interictal (seizure-free) and preictal (within 24 hours of seizure) network states. Our method focuses on connectivity patterns among a subset of informative iEEG electrodes, identified through permutation-based testing, enabling robust classification of brain states across time. We validate our approach using a leave-one-out cross-validation scheme and a pseudo-prospective forecasting strategy, assessing performance with metrics such as F1-score and balanced accuracy. Results show that low-dimensional Euclidean embeddings of iEEG connectivity yield interpretable and predictive markers of preictal activity, offering promising implications for real-time seizure forecasting and individualized therapeutic interventions. Journal: Physica A 1 Introduction The characterization of connectivity patterns in complex systems, from brain networks to large-scale infrastructures such as the Internet and social networks, exhibit intricate topologies that require sophisticated analytical frameworks capable of capturing their underlying structure and dynamics [ 3 ] . In the context of neuroscience, the study of intracranial electroencephalographic (iEEG) connectivity provides a crucial window into the functional organization of the brain, where network-based approaches have been employed to investigate seizure-related alterations in connectivity [ 26 , 7 ] . Various computational techniques, including coherence, mutual information, phase synchronization, and transfer entropy, have been applied to iEEG data to quantify neuronal interactions and detect patterns associated with epileptic activity [ 11 ] . A key aspect of seizure prediction research relies on comparing interictal (seizure-free) and preictal (seizure within the next 24h) states to identify early-warning signatures that may precede ictal events [ 2 , 21 ] . However, one of the major challenges in this field is selecting an appropriate mathematical framework to represent and analyze brain connectivity networks. Current approaches to predicting epileptic seizures largely rely on continuous EEG monitoring, using either scalp or intracranial recordings. This strategy demands substantial effort in both data acquisition and analysis. These models focus on detecting seizure-specific preictal patterns but are biased by changes related to shifts in vigilance states. As a result, they require either real-time identification of these states or the incorporation of diverse interictal reference periods. Recent studies have debated whether Euclidean or non-Euclidean (e.g. elliptic, Gaussian, hyperbolic) embeddings provide a more suitable representation for functional brain networks [ 1 , 4 ] . While some non-Euclidean (hyperbolic) embeddings have been assumed to be superior in capturing non-trivial (hierarchical or multi-scale) connectivity structures [ 1 ] , their optimization is prone to numerical difficulties [ 24 ] . Moreover, Euclidean embeddings offer a fundamental advantage: they provide an intuitive geometric representation with a well-structured distance metric in a space \mathbb{R}^{q} , facilitating a more interpretable analysis of embedded data (including the calculation of distances, barycenters and probability distributions). The use of latent geometric representations of anatomical brain networks allowed to identify and quantify the impact of epilepsy surgery on brain regions [ 23 , 19 ] . In a recent study, we have explored hyperbolic mapping of brain networks to identify discriminative features relevant to seizure prediction [ 18 ] . These findings highlight the necessity of integrating network-based approaches to track seizure-related alterations and optimize therapeutic strategies. Dimensionality reduction techniques play a pivotal role in network analysis, offering a statistical learning approach to condensing high-dimensional connectivity data while preserving essential structural and statistical information. This methodology applies to different brain network analyses, which typically follow a structured pipeline [ 25 ] : i) feature extraction: functional connectivity matrices are constructed using synchronization metrics such as the Phase Locking Value (PLV), mutual information, linear or non-linear correlation,…; ii) graph reconstruction: connectivity matrices are transformed into network structures, where nodes represent EEG channels and edges reflect their functional relationships; iii) dimensionality reduction: high-dimensional network data is projected onto a lower-dimensional representation that conveys, with minimal distortion, both local and global geometric information of the original connectivity graph; and iv) classification and prediction: machine learning models are trained to differentiate between different cognitive or neurological states. This study contributes to the field of computational neurosciences and epilepsy research by presenting a geometric framework that enables the iEEG data representation and classification. This approach strengthens the analytical toolkits for studying epileptic dynamics by enabling the forecasting of state-dependent features and network nodes with high discriminatory value. In contrast to continuous monitoring methods, our strategy aims to assess whether an individual’s iEEG connectivity patterns, recorded during short resting states with controlled vigilance, show differences between interictal (no seizure) and preictal (seizure within the next 24h) periods. We then explore the potential of these patterns to inform daily seizure risk predictions through the use of calibrated forecasting models. The structure of this article is as follows: we first describe the dataset and preprocessing steps, then detail the embedding and alignment methods, followed by the biomarker formulation and classification strategy. We conclude with a presentation of the classification and forecasting results, and a discussion of their implications in the broader context of network neuroscience and clinical applications. 2 Data acquisition The dataset consists of daily 10-minute resting-state intracranial EEG (iEEG) recordings from 10 patients (mean age: 30.7 years, mean duration of daily recordings: 11 days) with drug-resistant focal epilepsy, collected between January 2019 and July 2021 at the Epilepsy Unit of Pitié-Salpêtrière University Hospital (Paris, France). The study adhered to the Helsinki Declaration and was approved by an institutional review board (projects C11-16 and C19-55, sponsored by the National Institute of Health and Medical Research). Each 10-min resting-state period was labeled as “inter-ictal” if no seizure occurred in the next 24h ( N_{1}=69 ), or “preictal” if at least one electroclinical seizure occurred in the next 24h ( N_{2}=38 ). The implanted regions and the number of electrodes varied across patients, ranging from 10 to 62 electrodes (median = 20). For the purpose of the present study, only electrodes localized within the seizure onset zone (SOZ), as identified by the clinical team based on ictal activity, were retained for subsequent analyses. Details on the demographic information of each patient as age and number of recordings per day are available in the original study from this cohort [ 7 ] . 3 Brain networks construction Phase-locking value (PLV) was used to construct connectivity matrices from the iEEG data, which were computed between pairs of EEG signals during 20-second non-overlapping epochs, yielding 30 connectivity matrices per patient day. Synchrony values for the PLV were calculated for typical EEG frequency bands: delta ( \delta , 1–4 Hz), theta ( \theta , 4–8 Hz), alpha ( \alpha , 8–13 Hz), beta ( \beta , 13–30 Hz), low gamma (low \gamma , 30–49 Hz), and high gamma (high \gamma , 51–90 Hz). Only contacts located in the gray matter were considered, and a bipolar

Read the complete text