# Large Connectome Model: An fMRI Foundation Model of Brain Connectomes Empowered by Brain-Environment Interaction in Multitask Learning Landscape

**Authors:** Ziquan Wei, Tingting Dan, Guorong Wu  
**arXiv:** [2510.18910](https://arxiv.org/abs/2510.18910)  
**Date:** 2025-10-21  
**Summary page:** https://www.summarizepaper.com/en/paper/2510.18910v1/

## Key points

- Novel approach to developing a reliable foundation model of functional neuroimages by incorporating brain-environment interactions
- Focus on brain connectomes and their interaction with environmental variables in a multitask learning landscape
- Proposal of a scalable model architecture leveraging rich environmental variables and demographic data alongside functional neuroimages
- Multitask learning approach involving tokenizing multiple brain-environment interactions (BEI) for pretraining and semi-supervised finetuning
- Evaluation of the model across various applications such as sex prediction, human behavior recognition, and early diagnosis of Autism, Parkinson's disease, Alzheimer's disease, and Schizophrenia

## Summary

, , , , 

The study introduces a novel approach to developing a reliable foundation model of functional neuroimages by incorporating brain-environment interactions. This model focuses on brain connectomes and their interaction with environmental variables in a multitask learning landscape. The current limitations of AI models in clinical applications, due to small sample sizes, have prompted the use of self-supervised learning on extensive unlabeled fMRI data to pretrain large models. However, the disconnect between self-supervision and brain-to-outcome relationships often leads to suboptimal performance in downstream tasks like disease outcome prediction. To address this issue, the researchers propose a scalable model architecture that leverages rich environmental variables and demographic data alongside functional neuroimages. This multitask learning approach involves tokenizing multiple brain-environment interactions (BEI) for pretraining and semi-supervised finetuning by assigning pseudo-labels based on pretrained BEI. By incorporating diverse BEIs related to demographics and phenotypes, the model demonstrates scalability and outperforms classical regression models even without finetuning. The evaluation of this foundation model spans various applications such as sex prediction, human behavior recognition, and early diagnosis of Autism, Parkinson's disease, Alzheimer's disease, and Schizophrenia. The promising results suggest significant potential for enhancing current neuroimaging applications in clinical routines. Overall, the study highlights the importance of considering brain-environment interactions in developing robust models for functional neuroimaging analysis.