BHD2025 Projects

Take a look at BHD2025's projects!

Helmholz never published in Nature Neuroscience
(@qtabs)

Surprisal theory in psycholinguistics proposes that the harder a word is to predict from context, the more cognitive effort it takes to process. This project applies surprisal measures to characterize major journals in cognitive neuroscience and experimental psychology, testing whether surprisal distributions align with journal-level metrics (impact factor) or article-level metrics (citations). We will automate pipelines to extract articles, compute surprisal using open-source LLMs, run exploratory analyses, and write a report exploring whether highly cited papers in neuroscience embrace more or less linguistic surprise.

Goals

1. Set up an automated pipeline to scrape full-text neuroscience articles from 20-100 major journals in the field
2. Compute surprisal scores using an open-source LLM to calculate average word-level surprisal for a minimum of 50 articles per journal
3. Define exploratory analyses and collect appropriate journal/article metrics (e.g., impact factor, year, citations, sub-field)
4. Produce visualizations comparing surprisal distributions with the collected metrics
5. Generate collaborative interpretation of the results
6. Draft a short report summarizing context, methods, results, and limitations

Implementing denoising within phys2cvr software
(César Caballero Gaudes)

This project will aim to implement a denoising module within the phys2cvr software for cerebrovascular reactivity mapping. Briefly, we will create a module that implements the removal of the nuisance regressors (i.e. denoising) based on the lagged-GLM approach of phys2cvr. This will involve understanding the phys2cvr pipeline and developing an efficient approach of the denoising step.

Goals

The deliverable will be to implement this step and make it work.

FAIR-PsyFlow: making psychological assessment methods and data FAIR by design
(Inês Almeida)

This workshop aims to discuss barriers, opportunities and best practices in psychological assessment (with extensions, input from and output to behavioural research) using the lens of responsible and open science, reproducibility and FAIR principles. The end goal is to provide an opportunity for critically thinking about these topics, hands-on training, and to create a workflow to guide researchers in best practices from the project starting point.

Goals

The workshop aims to build insight and develop a best-practices workflow (in line with responsible and open research, and FAIR principles) for psychology that can be used, collaboratively worked and reused by others - with particular focus in the field of psychological assessment methods and data. The goal it to make methods selection, application & reporting, and data management & sharing more transparent, easy and best practices compliant, motivating all to prepare responsibly and share transparently their research practices and outputs – which are more diverse than a final thesis, research article or publication!
By the end of the workshop, a workflow (draft version) to implement in GitHub or in another collaborative platform should be created. This should contain key steps, best practices and useful resources to think critically when designing a study or project in Psychology research (in particular when using psychological assessment methods and data).

"I know that I know nothing", or do I? A workshop on priors for Bayesian Modeling
(Iñigo Urrestarazu-Porta)

Bayesian Data Analysis (BDA) offers a powerful toolkit to study human cognition and, overall, behaviour, where data collection tends to be very demanding, limited, and time consuming. Yet, BDA often becomes a daunting experience as it requires that the analyst specifies prior distributions for all model parameters, and newcomers to BDA often feel that they do not know at all how to define sensible priors for their analysis. In consequence, they rely on the default priors used by the software implementation used, e.g. brms (Bürkner 2017).
In the workshop, we will work on how to define (sensible) priors for a range of data sets. First, we shall see that, upon reflection, reasoning and the most simple knowledge of the world allows to constrain the parameter space of priors. Second, researchers are experts in their own fields, and we will discuss how we can leverage our expertise to find (more) suitable priors. Finally, we will approach prior definition through data simulation, as an empirical workflow to reassure our choice of prior distributions. Thus, the workshop aims to empower analysts to take full advantage of the possibilities offered by BDA.

Goals

The goal is pretty simple: learn how to think about priors, and apply that knowledge to your own research.

Developing a tool to estimate and report the carbon footprint of neuroimaging research
(Nick Souter)

This project will focus on workshopping, building, and disseminating a single tool which will allow neuroimaging researchers to estimate the environmental impact of their research, based on factors including imaging modality, duration of data collection, and analysis software choice. Such a tool could be used prospectively during grant writing, or retrospectively during manuscript writing.

Goals

1. Establish the scope of this tool including imaging modality (MRI only, or more diverse?), and aspect of the research process to be estimated (e.g., data collection, analysis, and dissemination)
2. Collate appropriate benchmark statistics relating to the carbon footprint of various aspects of the research process
3. Establish and the most appropriate format for this tool (e.g., Python package, dedicated webpage)
4. Start work on a digital tool to allow neuroimaging researchers to estimate the environmental impact of their work
5. Produce clear documentation to facilitate easy tool use and explain methodology
6. Produce branding for the tool, including a clear and effective name and appealing visuals

Open Data BIDSifier: LLM-based BIDSifier and metadata harmonization for non-BIDS datasets
(Stefan Dvoretskii)

The goal of this project is to explore if LLM-based workflow machines (sometimes calles "AI agents) can meaningfully assist in metadata harmonization, data format transformation and data preprocessing tasks that precede AI model inference or training. These tasks are time-consuming yet essential to reproducible, large-scale neuroimaging research.
While the BIDS standard ensures interoperability, there are some datasets for which no BIDS annotation is available. This is a "dead data" which can not be used on-par with BIDS datasets.
In this proof-of-concept, we aim to determine whether a coordinated system of AI agents can reliably execute these operations and produce AI-ready dataset collections, similar to those hosted on platforms like HuggingFace Datasets: OpenMind, with minimal human intervention.

Goals

If successful, Open Data BIDSifier will serve as a foundation for an AI agent that can identify and harmonize different datasets from open data, making sure these are immediately usable for machine learning and statistical analysis.
If successful, the developed data harmonizer tool will be published as a Python executable, as well as a HuggingFace space.

Fast Voxel-wise fMRI Correlation Mapping Using Phase Derivatives in LayNii IDA
(Omer Faruk Gulban)

Functional MRI analysis increasingly demands both higher spatial and higher temporal resolution: mesoscopic fMRI (submillimeter) reveals cortical microstructure and layer-dependent signals, while ultrafast sampling (tens of Hz) captures rapid hemodynamic and physiological dynamics. Extracting meaningful functional relationships at these scales requires voxel-wise correlation and connectivity computations across very large spatiotemporal arrays, an operation that is computationally intensive and often prohibitive with standard toolboxes.
Here we introduce LayNii IDA, a purpose-built, highly optimized implementation for voxel-wise correlation and related voxel-wise operations that (i) scales to submillimeter whole-brain or slab fMRI volumes, (ii) handles ultrafast single-slice and multi-slice acquisitions sampled at tens of Hz, and (iii) integrates preprocessing and on-the-fly statistical transforms to minimize I/O and memory overhead. We demonstrate that LayNii IDA enables exploratory, data-driven investigation of mesoscopic functional structure (layer- and column-scale), vascular structure, dynamic functional connectivity, and rapid quality control workflows that previously required ROI strategies. For this Brainhack, we specifically aim to extend IDA with phase-based fMRI time-series analysis using spatial phase derivatives. This implementation will allow voxel-wise correlation analyses in the phase domain, offering an alternative to the conventional magnitude domain fMRI time series.

Goals

1. Load and interact with phase fMRI time series data
2. Implement circular grayscale and/or variants to visualize phase timeseries within the GUI
3. Implement phase spatial derivatives (first and second order)
4. Explore various phase fMRI datasets using voxel-wise correlations
5. Report interesting findings.

Organizing Team

Brainhack Donosti 2025

Clementine Levy-Fidel

Chair

Monika Utrosa Skerjanec

Chair

Noemi Bonfiglio

Treasury

Joaquin Ordonez

Treasury

Jose Antonio Gonzalo Gimeno

Treasury

Melissa Donati

Website/Design

Marco Flores-Coronado

Website

Daniel Nieto Carrero

IT

Francisca Campos Matias

Program

Irtisha Chakraborty

Program

Cristina Comella Luengo

Program/Design

Chiara Bernini

Design/Public relations

Nirmitee Mulay

Design/Public relations

Antje Walter

Public relations

Seyma Takir

Website

Feel free to contact us!