Talks

Guillaume Dollé : Physiological pattern detection of neonatal EEG: a U-Net segmentation approach

The interpretation of EEG signals presents significant challenges due to substantial inter-subject variability, the influence of noise sources, including physiological artifacts (e.g. myogenic activity) and the fundamental non-stationary nature of neural oscillations.

Hypoxic-ischemic encephalopathy (HIE) is a recurrent perinatal condition where early diagnostic and intervention (e.g mild therapeutic hypothermia) is critical to mitigate neurological sequelae. Clinicians interpret the signal by identifying physiological and pathological patterns, that characterize the functional brain development of the patient.

We present our current work on a semi-synthetic data augmentation strategy for training convolutional neural networks (U-Net architecture) to segment non-stationary neurophysiological time series. By modeling patterns as stochastic perturbations of baseline EEG processes, we create semi-synthetic training data preserving the covariance structure of real signals while controlling the exploration of rare events. We evaluate the approach on neonatal HIE data, with particular attention to the bias-variance tradeoff in synthetic data generation, and the model's robustness to distributional mismatches between training and clinical data. We discuss our preliminary results and the underlying difficulties to this approach.

This work was supported by research grants ANR-22-CE45-0034 and ANR-19-CHIA-0015 and the American Memorial Hospital Fondation (AMHF).

Anna Dudek : Machine learning-based person identification using cyclostationary features of the electrocardiogram

We propose a novel approach for classifying electrocardiogram (ECG) signals. Specifically,
we model the data using an amplitude-modulated time-warped almost cyclostationary process
and extract relevant features for various machine learning algorithms. In particular, we focus on the
Fourier coefficients of the autocovariance function of the underlying cyclostationary process. We compare our results with those obtained from the same machine learning algorithms, but using the magnitude of the Fourier coefficients from raw ECG signals. Our experiments validate the effectiveness of our approach for biometric purposes, i.e., for person identification on the basis of their ECG signal classification and demonstrate the benefits of using features derived from the underlying cyclostationary signal of each ECG signal.

Thierry Dumont : Adaptive EEG Segmentation via Composite-Entropy Clustering: Artifact Cleaning and Regime-Change Detection

We present a framework for EEG segmentation based on a composite-entropy minimization criterion with data-driven model order given by the Relative Entropic Order (REO). The criterion balances mixture-weight entropy with cross-entropy to a reference family, yielding hard partitions and consistent estimates of the effective number of clusters. Its link to Classification-EM (CEM) clarifies component pruning and provides a monotone-descent surrogate objective.
Building on these results, we implement a windowed multichannel pipeline that isolates electrical   and detects regime changes via shifts in channel means, variances, and correlations. On real recordings, the method separates artifacts from neural activity and flags state transitions with little tuning. This entropy-based formulation offers a transparent, theoretically grounded approach to EEG cleaning and change-point detection.
 

Jean Marc Freyermuth  : Harmonizable time series in EEG data analysis.

Harmonizable time series extend the concept of stationary time series by allowing a spectral decomposition in which the components are correlated. As a result, the covariance function of a harmonizable time series is bivariate and admits a two dimensional Fourier decomposition, known as the Lo`eve spectrum. In this talk, we introduce a parametric form for harmonizable processes, specifically Harmonizable Vector AutoRegressive and Moving Average models (HVARMA). We present a method for generating finite time sample realizations of HVARMA with known Loève spectrum. Finally, after discussing a nonparametric approach to estimate the spectral characteristics of spatiotemporal processes that exhibit local time harmonizability, we illustrate how harmonizable processes can aid in analyzing functional connectivity in EEG data.

Theo Gnassounou : Spatio-Temporal Normalization for learning in biosignal

Distribution shift poses a significant challenge in machine learning, particularly in biomedical applications using data collected across different subjects, institutions, and recording devices, such as sleep data. Existing normalization procedures of the data struggle to completely help mitigate distribution shifts.
When applied over the time dimension, they ignore the dependencies and autocorrelation inherent to the vector coefficients they normalize. This presentation will introduce an Optimal Transport (OT) based method that adapts the cross-power spectrum density (cross-PSD) of multivariate signals by mapping them to the Wasserstein barycenter of different domains.

Claudia Kirch : Change points in time series and data segmentation algorithms

Many statistical and AI procedures rely on the assumption that the stochastic nature of the underlying data remains constant over time while in practice this behaviour will often change, a phenomenon known as concept drift in machine learning. If such changes are not taken into account the subsequent analysis does not yield reliable results. 
In this talk, we will give an introduction into the field of change point analysis, featuring also some examples from neuroscience. Then, we discuss data segmentation algorithms based on moving sum procedures, that estimate the number and locations of multiple changes in an efficient way. This includes recent work on general methodology that can be used in a variety of situations, from robust methodology for the detection of changes in the mean to the detection of changes in linear and non-linear time series including count time series. The talk will give a survey over several papers joint with various co-authors.

Frédéric Morain-Nicolier : EEG Signal Characterization and Classification with Time-Frequency Distributions

This presentation will focus on exposing a general methodology from classifiers, Time–Frequency Distributions, and dissimilarity measures for epileptic seizures detection.
Based on a published paper in 2021, these points will be discussed :
- Characterizing the evolution of the frequency content of a non-stationary signal as a function of time : Wavelet Transform vs Time–frequency representations
- How to compute dissimilarities between EEG TFD and classify signals
- Application and examples in Epileptic and neonatal EEG.

Hernando Ombao : Modeling Frequency-Specific Lead-Lag Associations: With Applications to Rat Brain Waves

This talk will investigate the phenomenon where brain signals from two regions exhibit a lead-lag relationship that varies depending on the frequency of the underlying oscillatory activity. This is exhibited in cognitive tasks where low-frequency oscillations might be contemporaneous but, for higher-frequency oscillations, one region may lead another region. This is not covered under the usual setting where the lag between regions is constant across frequencies. We will develop a novel general model based on the Cramer representation but with frequency-specific phases. We will derive the stochastic properties of this model and, in particular the phase function which could be non-linear. To estimate the frequency-specific lead-lag between time series, our approach is to select the best piecewise-linear approximation to the true phase function which is estimated via change-point algorithms. Using this proposed method, we investigate the neuronal lead-lag relationship between local field potentials recorded from rats in an olfaction memory experiment. This is joint work with Sipan Aslan (KAUST), Paolo Redondo (KAUST) and Adam Sykulski (Imperial College).

Efthymios Papatzikis : EEG (Pre-)Processing and Quantitative Analysis in Neonatal Research: A specialized pipeline

Traditional clinical EEG, such as amplitude-integrated EEG (aEEG), while useful, offers limited insight into the intricate brain dynamics of newborns, and most importantly preterm infants, due to its broad analytical scope. This limitation highlights a critical research gap: the absence of a specialized, quantitative EEG (qEEG) processing and analysis framework designed specifically for neonates undergoing neurodevelopmental treatments and interventions. Such a framework is essential for delving into the nuanced effects of neurosensory related tools on the developing brain, providing a deeper understanding of how these stimuli influence neurodevelopmental trajectories. 

Addressing this gap, this presentation will discuss the development of an advanced qEEG analysis pipeline tailored for the neonatal population in NICUs. By leveraging refined data, preprocessing/processing techniques and analytical methods suitable for the unique characteristics of neonatal EEG data collected during a music-based intervention, this pipeline aims to offer new insights into the neurophysiological responses of preterm infants. The outcome of this approach could significantly enhance our understanding of infant brain development, paving the way for optimizing neurosensory interventions in neonatal care.