Modeling Neural Activity With Transformers To Predict Impulsivity

Sophie Quynn, Saeedeh Komijani, Samuel M. McClure, Julie B. Schweitzer, Dipak Ghosal

Keywords: ADHD, Impulsivity, Reward Processing, Neural Basis, fMRI, Machine Learning, Transformers, Attention Maps, Explainability, Classification

Abstract: Impulsivity is a key behavioral concern associated with numerous mental health disorders and attention-deficit/hyperactivity disorder (ADHD) in particular. Assessment of impulsivity traditionally relies on rating scales, particularly in the clinical setting. These measures have known limitations due to their subjective nature, including potential to be affected by recency effects, cultural bias, contextual factors, and their inability to assess underlying cognitive processes. As a consequence, there has been a long-standing effort to identify the biological basis of impulsivity, using neuroimaging techniques such as functional MRI (fMRI). We propose a machine learning approach that integrates behavioral measures of impulsivity and reward sensitivity with task-based fMRI data to identify patterns of brain activation associated with impulsivity in a group highly enriched with impulsivity and ADHD. Using a Win/Loss reward-processing task, we extracted regression coefficients (beta values) from a Generalized Linear Model applied to fMRI time series within Regions of Interest (ROIs) selected based on prior meta-analysis. The beta values, combined with spatiotemporal embeddings, were input to a transformer encoder to learn latent representations associated with high versus low impulsivity. We trained and validated the model, and further examined its internal reasoning using attention maps. The model achieves effective classification accuracy in distinguishing low versus high impulsivity across different subject groups, including adolescents and young adults—ages typically associated with greater impulsiveness and risk-taking. Furthermore, the attention maps show correspondence to the current understanding of the neural basis of reward processing and impulsivity in key ROIs. This work demonstrates the feasibility of applying transformers to fMRI-based tasks and is a promising tool to identifying patterns of brain activity associated with complex behavioral constructs with clinical importance.

Primary Area: applications to neuroscience & cognitive science

Submission Number: 1777