Researchers focus too much on statistical significance and have really bad intuitions about the practical significance of research findings. That is why the decades of neuroscience research has had so little impact on mental health care, in my humble opinion. To address this, I have developed E2P Simulator: an interactive, open-source, web-based tool that translates standard effect sizes into predictive utility metrics. It visually and quantitatively demonstrates the relationship between commonly reported statistical metrics (such as Cohen’s d and Pearson’s r) and predictive metrics (such as AUC. PR-AUC and Net Benefit), while accounting for real-world factors like measurement reliability and outcome base rates. Much like how power analysis tools (such as G*Power) help researchers plan for statistical significance, I have developed E2P Simulator, which plan for practical significance.

This Traintrack is a free-form, hands-on walkthrough demonstrating how various neuroimaging file formats can be loaded and visualized using Python scripts in a Jupyter Notebook. No prior software installation is required; participants can follow along via a Google Colab session. Examples will mostly follow along this Jupyter Book chapter.

Our science might be open, but maybe not as reproducible as would like. In between many different reasons, one of them is that the incentive to create reproducible deliverables is still low. To overcome that, I'd like to propose to the OSSIG and to OHBM leadership a reproducibility challenge that forces people to create FAIR deliverables. If you'd like to help me think about proposal, please join me for a chat!

Original proposal: challenge2pgs.pdf

This TrainTrack will focus on hands-on exploration of neural time series data—such as in vitro spiking activity from closed-loop neuronal cultures and fMRI recordings from human subjects—using cutting-edge machine learning methods. Participants will work on a project aimed at modeling these data using cutting-edge transformer-based models and graph neural networks.

Emphasis will be placed on predictive modeling, where participants will work on building systems that forecast future neural dynamics across different biological systems. In parallel, the session will also explore causal inference, integrating techniques like attention-based mechanisms and graph transformers to identify directed interactions and underlying causal structure in neural networks.

These two goals—predicting future neural dynamics and inferring underlying causal mechanisms—are deeply intertwined. Many of the same architectures, such as graph-based transformers, can be leveraged to both forecast neural activity and uncover the directed relationships that drive it. This session invites participants to explore models that unify both perspectives, enabling more interpretable and generalizable tools for neuroscience.