Neuroscience

Predicting mental states from functional Magnetic Resonance Imaging (fMRI) data through whole-brain decoders is becoming increasingly popular. To create interpretable decoders, choosing a level of parsimony that determines which brain signals should be included is necessary. The traditional approach has been to test the predictive performance of decoders with different levels of parsimony on new data through cross-validation. However, as fMRI data is rich in variables, the classic cross-validation (CV) approach often leads to inflated whole-brain decoders that include many random voxels. A classic correction approach is the 1 Standard Error (1SE) rule that accepts the most parsimonious decoder with a performance within 1SE of the best-performing decoder's performance in cross-validation. This approach ensures the choice of a more parsimonious decoder. However, it depends on the outlier-vulnerable metrics of the means from the CV performances and the standard error. To deal with this, we propose a pairwise fold (i.e. subject) comparison between the best-performing decoder from the CV and the more parsimonious decoders. The most parsimonious decoder with an insignificant paired t-test against the best-performing decoder is chosen. We show that the pairwise comparison approach is more outlier resistant than 1SE and excludes irrelevant voxels more reliably than the classic CV approach.

This is an in-person presentation on July 22, 2024 (10:00 ~ 10:20 CEST).

In the analysis of event-related functional magnetic resonance imaging (fMRI) data, inference on the brain's reaction to a presented stimulus is done with the help of the hemodynamic response function (HRF). It is common practice to assume that the brain's reaction is stationary, that is, the shape of the HRF corresponding to one type of stimulus does not change over time. However, this is not necessarily true; possible sources of nonstationarity are changes in emotions, stress level or learning effects. In this work we aim at answering the following questions: Can we assume stationarity of the hemodynamic response? If not, how does the HRF's shape change over time? In other words, we investigate if current methods sufficiently account for nonstationarity of the HRF. Subsequently, we present a methodology that allows for variation of the HRF's shape over time. To this end, we employ regression models that allow for time-varying beta coefficients to model the HRF's shape. In contrast to existing methods, we analyze fMRI data of multiple subjects, consider multiple regions in the brain and allow for variation in multiple shape parameters of the HRF. Therefore, the proposed procedure accounts for several sources of multiplicity. The procedure is applied to fMRI data of a categorization learning experiment to investigate changes in the HRF's shape over the course of learning.

This is an in-person presentation on July 22, 2024 (10:20 ~ 10:40 CEST).

Using statistical models for the analysis of neurophysiological time-series, such as electroencephalography (EEG) or pupil dilation recordings, is complicated by the fact that these signals change non-linearly over time. Additionally, other experimental continuous variables might have a non-linear effect on the measured signal as well. The analysis is typically further complicated by substantial between-subject and between-trial heterogeneity of these non-linear relationships. Generalized additive models (GAMs) are theoretically well-equipped to address both challenges, allowing the estimation of non-linear functions of predictor variables as well as random effects to account for these sources of heterogeneity. However, in practice it is often computationally intractable to include sufficient random effects, as it is not uncommon for cognitive experiments to involve thousands of trials across participants. Here, we combined and extended recently proposed strategies to reduce memory requirements and matrix infill into a sparse GAM estimation algorithm capable of handling previously impossible (non-linear) random effect structures. This allowed us to compute proper GAM models of pupil dilation data with ~1.8 million observations and EEG data with ~23 million observations. Fitting these models introduces new challenges for established model comparison strategies, which we investigated with simulation studies. Based on the results we established guidelines on how to identify the optimal model. To further facilitate this model-based analysis approach, we provide an openly available Python package of the algorithm and the investigated model comparison strategies.

This is an in-person presentation on July 22, 2024 (10:40 ~ 11:00 CEST).

Electroencephalography (EEG) is a fundamental tool in neuroscience, offering key insights into the complex workings of the brain. This study introduces a global model for EEG analysis based on a stochastic autoregressive framework derived from established models of neural behavior. While it is typically thought that EEG frequency bands emerge from synchronous synaptic activity, the global model of EEG states that delays in axonal propagation across corticocortical and thalamocortical connections significantly contribute to the variance observed in EEG signals. The present model predicts that spectral peaks in scalp-recorded EEG data can be solely attributed to axonal time delays at various distances. The autoregressive models are notable for their linear structure that efficiently captures temporal relationships within EEG signals, highlighting the impact of axonal propagation delays with greater computational efficiency. The model employs a connectivity atlas to determine the connectivity and distances between various brain regions. Additionally, it incorporates distributions of axonal delays and Event-Related Potentials (ERPs) in response to visual stimuli. The approach allows for an accurate reproduction of EEG power spectra, including both resting-state alpha rhythms and ERP peaks. The findings suggest that axonal delay times and neural connectivity within linear predictive models influence EEG dynamics, offering a method to analyze individual cognitive variations through EEG data. In the future, we aim to apply these models alongside cognitive frameworks to draw inferences about individual variations in neurocognition.

This is an in-person presentation on July 22, 2024 (11:00 ~ 11:20 CEST).