Deep learning models can predict brain disorders from functional MRI scans, but do they find the right brain regions? A new framework called RE-CONFIRM tests whether the biomarkers identified by these models are reliable. Researchers found that simply fine-tuning foundation models may overlook critical brain hubs linked to autism and ADHD.
The research
Scientists from Nanyang Technological University led by Deepank Girish and Yi Hao Chan developed RE-CONFIRM, a framework that evaluates how robustly deep learning models identify potential biomarkers of neurological disorders. They tested five large datasets covering Autism Spectrum Disorder (ASD), Attention-deficit Hyperactivity Disorder (ADHD), and Alzheimer's Disease (AD). Their results showed that standard performance metrics like accuracy do not guarantee that a model's highlighted brain regions are biologically meaningful. For example, fine-tuning a foundation model on ADHD data failed to consistently highlight the default mode network, a hub known to be involved in the disorder. To address this, they introduced Hub-LoRA, a fine-tuning method that helps models learn neurobiologically faithful biomarkers validated by meta-analyses. Hub-LoRA outperformed custom deep learning models and produced more trustworthy brain maps.
Why it matters
For anyone interested in brain health, this research highlights a crucial gap: not all AI findings in neuroscience are equally reliable. As AI tools become more common in diagnosing conditions like autism, ADHD, and Alzheimer's, patients and clinicians should question whether the brain signatures they highlight are real or just statistical artifacts. RE-CONFIRM offers a way to verify discoveries, making future brain-based diagnostics more trustworthy. For brain training enthusiasts, understanding which brain networks truly underlie cognitive skills can help focus training on regions that matter most.
What you can do
If you use brain training apps or read about brain biomarkers, stay curious about how AI findings are validated. Look for studies that use robustness checks similar to RE-CONFIRM. Meanwhile, practice cognitive skills like attention and memory—they strengthen real brain networks such as the frontoparietal and default mode networks.
Source: arXiv q-bio.NC
Curious about your own brain? Take our free adaptive IQ test or try 306 brain training levels.
