Foundation models for discovering robust biomarkers of neurological disorders from dynamic functional connectivity

A team of researchers from MIT and NTU Singapore has introduced RE-CONFIRM, a comprehensive framework for evaluating the robustness of potential biomarkers identified by deep learning (DL) models, including brain foundation models (FMs). In experiments across five large datasets covering Autism Spectrum Disorder (ASD), Attention-deficit Hyperactivity Disorder (ADHD), and Alzheimer's Disease (AD), the team found that standard performance metrics like accuracy are insufficient for assessing whether the biomarkers a model identifies are truly reliable. Simply fine-tuning FMs often fails to capture regional hubs known to be implicated in disorders like ASD and ADHD, leading to biologically inconsistent results.

To address this, the researchers developed Hub-LoRA (Low-Rank Adaptation), a fine-tuning technique that enables FMs to not only outperform custom DL models but also produce neurobiologically faithful biomarkers supported by meta-analyses. Hub-LoRA preserves the model's ability to focus on critical brain regions while adapting to new tasks. The RE-CONFIRM framework is generalizable and can be easily applied to any DL model trained on functional MRI datasets. Code is available on GitHub, opening the door for more robust and clinically useful biomarker discovery in neurological disorders.