Research projects

018. The BrainBeats toolbox (2024)

I developed the BrainBeats toolbox, implemented as an open-source EEGLAB plugin to facilitate the study of brain-heart interplay, using EEG and cardiovascular signals (ECG/PPG).

4 methods are available: 1) heartbeats-evoked potentials (HEP) and oscillations (HEO), which captures how the brain responds to heartbeats with millisecond accuracy; 2) Extraction of EEG and HRV metrics (time, frequency, and nonlinear domains) to find associations or differences between these features or or assessing Pre/Post-intervention changes; 3) Extract cardiac field artifacts (CFAs) from EEG signals with independent component analysis (ICA); and 4) calculate brain-heart coherence and causal interactions to better understand the interplay between the two systems. The toolbox is for both novices (general user interface) and experts (command line) and comes with a step-by-step tutorial.

The tooblox is open-source and available here

017. Scoring human creativity with AI (2024)

In this study, we evaluated the performance of an AI algorithm, the Open Creativity Scoring with Artificial Intelligence (OCSAI) system at scoring the Alternate Use Task (AUT), which tests some aspects of human creativity. We saw a strong correlation between manual (3 human raters per rating) and OCSAI scores for elaboration (rho = 0.76, p < 0.001, n = 520), indicating that the AI-based system effectively captures the elaboration component of creativity (i.e., the level of detail and development in responses). The correlation for originality component was weaker but significant (rho = 0.21, p < 0.001, n = 520), with originality referring to the uniqueness and novelty of ideas. These findings highlight the potential of AI-driven methods like OCSAI to automate the scoring for this task, especially for elaboration, while suggesting that further refinement may be needed for accurately assessing originality.

016. Investigating a unique visual experience (2024)

This case study examined a unique visual phenomenon, the visual perception of vivid holographic images overlaid on their visual field.

Utilizing state-of-the-art source reconstruction methods from EEG signals (64 electrodes), the study identified the brain regions involved in Upsight and assessed functional connectivity to determine how these regions or networks communicated, we observed a large decrease in alpha power (11 Hz). The findings suggest that Upsight shares more similarities with normal vision than with VMI, imagination, or visual hallucinations.

015. EEG and HRV correlates of well-being using low-cost wearable neurotech (2024)

Wearable devices that measure brain waves (EEG) and heart activity (ECG) could be a simple and low-cost way to check how well people are feeling in everyday situations. However, it's tricky to get clear signals and find reliable indicators of well-being. We tested if these devices could identify markers of well-being by collecting short EEG and ECG recordings from 60 people in real-world settings. We looked for patterns in the data that might relate to different aspects of well-being, like happiness, life satisfaction, physical health, and social connections. Our study found some links between heart activity and psychological well-being, and between brain activity and physical well-being. These findings suggest that wearable devices could offer a quick and easy way to monitor well-being in daily life.

014. Source analysis of an unusual visual experience (2024)

An individual reports perceiving "holographic" images as though they appear on an inset screen that overlays with the ordinary visual field. This case study compared the underlying brain electrical activity with a control condition of visual imagery. Robust methods were employed using LCMV source localization and reconstruction, mass-univariate nonparametric statistics, and robust corrections for multiple comparisons. Large reductions in alpha and beta spectral power were observed during the experience relative to control at scalp and source levels, with a peak at 11 Hz. Interestingly, the effect was lateralized to the left frontoparietal areas at the source level, interpreted in terms of the asymetric inhibition model (see Grimshaw and Carmel, 2014). That is, we suspect the effect to reflect stronger executive control during the control condition to block the unusual visual stream to successfully recall the stimulus via conventional visual imagery.

013. Classifying EEG signals with machine learning (2023)

This project aimed to train a machine learning (ML) model to classify EEG signals (good/bad) collected with the MUSE headsets. First, I manually labeld 3000 30-s segments and extracted features in the time, frequency, and nonlinear domains. Then, I trained an ensemble of ML models (decision trees, logistic regression, LDA, SVM, Naive Bayes, neural networks) implementing feature-selection, PCA-dimension reduction, hyperparameter tuning, and 5-fold cross-validation (on 80% of the data). The best model was then validated on a test dataset (20% remaining data from different subjects to avoid overfitting). The best models achieved 93.5% of accuracy for frontal channels and 91.4% for the posterior channels. These classifiers can be used via my import_muse plugin to classify bad MUSE channels during importation.

012. Advanced re-referencing methods for EEG data (2021)

I make two advanced re-referencing methods for multidimensional EEG data: the reference electrode standardization technique (REST; Yao 2001) and the Surface Laplacian transformation (also called current-source density transformation). The image illustrates the statistically significant differences at the grand average ERP level (N = 81, 230 trials each, after spatiotemporal cluster correction, p = 0.01), comparing the brain's response to emotional stimuli (pleasant and unpleasant valence with high arousal) relative to neutral ones (no emotional valence and low arousal). We can see that the reference to infinity (REST) is preferrable to the commonly-used reference to average when assessing global (widespread) scalp dynamics, whereas the Surface Laplacian is best to capture local dynamics, dealing better with undesired volume conduction effects (especially when for source analyses).

Both are accessible via command line for easy implementation and automation.

011. 3D ERP video (2021)

3D visualization of the spatial distribution of an averaged event-related potential (ERP) over several hundred milliseconds from -100 to 450 ms after stimulus presentation (at time = 0). Current-source density (CSD; i.e. Surface Laplacian) transformation was performed on this subject's 64-channel EEG data to increase spatial resolution and local dynamics.

010. Brain predictive processes (ongoing)

The contingent negative variation (CNV) is a well-known ERP component of brain anticipation, depicted as a slow negative deviation in frontocentral areas. The CNV reflects the capacity of the brain to learn the temporal regularities (the "internal clock"), independently of motor preparation. However, less is known about what the brain does when there are no regularities (i.e. unpredictable settings). In this study implementing state-of-the art methods (latest signal processing tools, hierarchichal linear modeling, robust statistics, N = 81, etc.), we found that the brain may be able to anticipate forthcoming emotional stimuli relative to neutral ones despite seemingly unpredictable conditions.

009. Avoiding EEG "Ghost ICs" (2023)

Many EEG researchs use independent component analysis (ICA) to process and anlayze their EEG data. However, some common processing steps (e.g., electrode interpolation or re-reference to average) can lead to effective rank defficieny, which in turn, leads to the emergence of "ghost ICs". These independent components display a typical scalp topography, but actually contain white noise properties in both time and frequency domains. They can therefore be easily missed and significantly affect findings in unknown ways. See the full-article here for more details.

008. Importing EEG data (2022)

I developed the import_muse() and import_edf() tools to import data files from various formats (.csv, .edf, .edf+, etc.) into EEGLAB. Impleted as EEGLAB plugins, they can be used with the GUI or command line. The plugins detect the data types (e.g., EEG, PPG, GYR, ACC, AUX, etc.), the sample rate, annotations if any, etc. and convert everything to the EEGLAB format. Step-by-step tutorials are available.

007. Extracting entropy measures from biosignals (2022)

I developped the get_entropy() toolbox, implemented as an open-source EEGLAB plugin allowing to compute entropy measures from biosignals (e.g., EEG, MEG, ECG). Nonlinear measures are particularly promising for capturing ocmplex, multiscale (time and space) dynamics that may be missed by conventional measures (time or frequency analysis). The plugin currently supports: sample, approximate, and fuzzy entropy (uniscale); multiscale entropy (MSE), multiscale fuzzy entropy (MFE), and refined composite multiscale fuzzy entropy (RCMFE). A bandpass filter can be applied at each scale factor to reduce spectral contaminations. Since these entropy measures are very computation-heavy, the plugin implements GPU and parrallel computing when possible.

006. Validating EEG signals from a wearable system (2021)

This project aimed to assess whether the Muse (InteraXon, Inc.), a wearable EEG headset, can be used to reliably capture EEG signals and relevant measures (alpha asymmetry and individual alpha frequency). A minimal amount of data was deliberately collected to test the feasibility for real-world applications (EEG setup and data collection being completed in under 5 min). The MUSE's temporoparietal (TP) channels showed comparable power spectra, alpha asymmetry, and individual alpha frequency (IAF) relative to that obtained from a gel-based state-of-the-art BIOSEMI system (referenced to both Fpz like the Muse, and to average, the gold standard). However, the frontal channels needed to be re-referenced to linked-mastoids to provide valid outputs. Furthermore, we observed satisfying internal consistency reliability. See publication here for more details.

Mastoid-ref montage (frontal channels)

Fpz-ref montage (TP channels)

005. Identifying the main well-being dimensions (2021)

This project aimed to identify the main dimensions of multidimensional well-being (MWB). 2647 individuals participated in an online survey between November 19, 2020, and September 26, 2021, during the COVID-19 pandemi. 1615 were unique pre-intervention records and 429 were unique post-intervention. After validating a quick visual analog scale for capturing multidimeniosnal well-being in a single question (convergent and test-retest validity), a multiple regression model showed that the main dimensions of MWB were the hedonic WB (i.e., affective WB), eudaimonic (psychological WB), physical WB (i.e., health and pain levels), and social (sense of connectedness with others) dimensions. The model explained 44.6% of the variance in MWB (N = 1443).

Multidimensional well-being was also influenced by reported connection with nature, religion/spirituality, physical activity during leisure, and personality trait (N = 415). Physical activity at work, meditation practice, relationship status, and creativity levels did not show any relationship with MWB.

004. Self-health monitoring and wearable neurotechnology (2020)

BCIs and wearable neurotechnologies are being used to collect real-time neural and physiological data, showing great promise for advancing medical diagnostics, prevention, and intervention. This book chapter discussed here presents wearable electroencephalography systems, highlighting their groundbreaking innovations in real-time health monitoring and their technical pros and cons. Though these systems offer potential for large-scale data collection beyond traditional laboratory settings, they also face methodological and ethical challenges that must be addressed in the current and future research.

003. Physiological examination of an altered state of consciousness (2019 & 2023)

Some individuals are able to enter deep trance states at will, where their consciousness is highly altered (no awareness, no recollection, feeling loss of control over the body, etc.). We assessed how physiological activity (64-channel EEG, ECG, GSR, voice) would differ between this state an a baseline (directed mind wandering).

A 1st anlaysis only showed differences in the voice data. A 2nd analysis addressed some limitations and showd increased EEG delta and theta power in the frontal region and increased gamma in the centro-parietal region during mind-wandering, whereas trance showed increased beta and gamma power in the frontal region. A source anlaysis (brain areas spectral power and functional connectivity) showed no differences. However, reported trance depth was negativley correlated with whole-brain connectivity in all frequency bands suggesting deeper trance was associated with less overall brain functional connectivity

002. Brain oscillations in color vision (2016)

I did my 2nd Master's project with Dr. Rufin VanRullen. In a previous study, Rufin identified the brain's preferred response to luminance, the so-called "echo function" (an EEG oscillation displaying reverberations at 10 Hz up to 1 s after the presentation of the stimulus; see VanRullen & MacDonald (2012) ). The goal of my Master's project was to determine whether this echo functions was also the brain's preferred way of processing color stimuli (namely red, green, blue, yellow). The project involved developing psychophysics code to find individualized equiluminance threshold, modifying stimulus presentation scripts, and recording and analyzing 64-channel EEG on 12 subjects.

Results suggested that no equivalent echo function exists for color vision.

001. Brain atrophy in Schizophrenia (2015)

During my first research project with Dr. Sonia Dollfus (1st year of MS.c.) at the ISTS lab in Caen, France, I learned how to process and analyze MRI data to measure grey matter atrophy in patients with Schizophrenia.

3T MRI scans were segmented and normalized using DARTEL in SPM12, allowing tailored template of the study population, reducing normalizing errors and enhancing alignment of small structures. After smoothing, intracranial volume was calculated using GM, WM, and CSF. A voxel-based morphometry (VBM) analysis (voxel-by-voxel comparisons) was conducted to characterize GM differences between our two groups (63 controls vs 51 patients), accounting for age, gender, education, and intracranial volume as covariables.

Whole-brain analysis showed atrophies consistent with the literature (median orbitorontal, bilateral insula, and left middle cingulate cortices; p-corrected = 0.001).

Then, the contrast map was projected onto a 3D mesh of the hippocampus and its subfields. This mesh was manually delineated on coronal slices by Renaud LaJoie (see publication here ). In the right hippocampus, atrophy was detected in the median-posterior part of the Subiculum and other subfields, namely CA2-3 and GD-CA4, as seen in the dorsal view. This extended across almost the entirety of the CA1 region and marginally into the ventral Subiculum. In contrast, the left hippocampus showed a milder atrophy primarily concentrated in the posterior section of the Subiculum, discernible both ventrally and slightly dorsally. However, these differences did not survive the correction for FWE.