Summary: For years, neuroscientists have prioritized the strongest 10% of brain connections in imaging studies and treated the remaining 90% as background “noise.” A recent study from Yale shows that those weaker connections—the ones usually discarded—can predict behavior just as well as, and sometimes better than, the strongest signals.
The study indicates that predictive information is widely distributed across the brain. Rather than a single, definitive network underlying a behavior, multiple distinct networks can carry equally valid predictive power.
Key Findings
- Multiple Pathways: Researchers found numerous, non-overlapping networks capable of predicting the same behavior, highlighting redundancy and functional flexibility in brain organization.
- Psychiatric Implications: For disorders such as depression, different individuals may depend on entirely different neural routes to produce the same symptoms or behaviors.
- Therapeutic Targets: If diverse circuits can predict a condition, treatments should not focus only on the most prominent networks. Overlooked circuits might provide new options for patients who do not respond to standard interventions.
- The Accuracy Myth: Strong statistical association does not necessarily equate to exclusive biological relevance. Signals once dismissed as noise may prove essential for more personalized diagnostics and therapies.
Source: Yale
Many imaging studies simplify brain data by keeping only the strongest signals, but that approach risks missing important information.
A new paper in Nature Human Behavior shows that connections commonly excluded during feature selection can, in aggregate, predict behavioral outcomes with high accuracy and point to different neurobiological explanations. The implications reach from basic research to clinical practice.
“Feature selection simplifies data by focusing on a small subset of connections, but that can hide other meaningful patterns,” says Brendan Adkinson, PhD, the study’s lead author and an MD-PhD student at Yale School of Medicine. “Our results suggest that multiple, non-overlapping networks can predict the same behavior just as well.”
Overlooked brain connections
A major challenge in human neuroimaging is the sheer scale of connectivity data. To make models interpretable, many studies use feature selection that emphasizes the top 10% of connections. To test what gets lost in that process, the researchers analyzed brain imaging and behavioral data from more than 12,000 participants drawn from four large U.S. datasets.
For each participant the team computed how strongly each brain connection related to the outcome of interest. They then ranked connections from strongest to weakest and split them into ten equal, non-overlapping groups. Group one contained the usual top 10% of features; groups two through ten comprised the remaining 90% typically treated as noise. The researchers built ten separate prediction models, one per group.
Surprisingly, models trained on lower-ranked groups—groups two through nine—often matched the prediction accuracy of the top group. In some cases, these lower-ranked models outperformed the top-10% models. The authors interpret this as evidence that predictive information is distributed widely across many connections, not concentrated solely in the strongest signals.
“Even when we excluded the networks researchers normally rely on, we still reached nearly the same accuracy using what’s typically discarded,” says Adkinson, who works in the lab of senior author Dustin Scheinost, PhD, associate professor of radiology and biomedical imaging at Yale School of Medicine.
Individual differences in mental health
Focusing narrowly on the strongest features risks oversimplifying brain–behavior relationships, particularly in psychiatric conditions. For example, people with depression may express similar symptoms via different neural pathways. If multiple circuits can predict the same outcome, clinical interventions should consider a broader range of targets.
“The networks targeted by current treatments may help many patients, but other, overlooked networks could be crucial for individuals who don’t respond,” Adkinson explains. Incorporating broader patterns into brain-based biomarkers may improve personalized care and the reliability of findings across studies.
Key Questions Answered:
A: Practical constraints. The brain has millions to billions of possible connections, and researchers simplify analyses by selecting the most strongly associated features. This study shows that subtler, distributed signals can carry similar information and deserve attention.
A: Not necessarily wrong, but incomplete. Existing therapies often target prominent networks, which benefit many patients. This work suggests additional networks might be more relevant for people who do not respond to those interventions.
A: Potentially yes. Expanding models to include distributed, weaker signals could yield richer, more precise biomarkers. Instead of searching for a single signature of a disorder, clinicians could identify which pathways are most relevant for each person.
Editorial Notes:
- This article was edited by a Neuroscience News editor.
- The journal paper was reviewed in full by editorial staff.
- Additional context and explanation were added by the editorial team to clarify implications for research and clinical practice.
About this mental health and neuroscience research news
Author: Colleen Moriarty
Source: Yale
Contact: Colleen Moriarty – Yale
Image: The image is credited to Neuroscience News
Original Research: Open access.
“Feature selection leads to divergent neurobiological interpretations of brain-based machine learning biomarkers” by Brendan D. Adkinson et al., Nature Human Behavior. DOI: 10.1038/s41562-026-02447-y
Abstract
Feature selection leads to divergent neurobiological interpretations of brain-based machine learning biomarkers
Understanding the neurobiology behind cognition and mental health is a central aim of human neuroimaging. Machine learning models trained on imaging data are increasingly used to predict behavior and enhance precision medicine. Yet the high dimensionality of connectivity data makes interpretation difficult, so researchers commonly use feature selection to simplify models.
This study used four large datasets totaling over 12,000 participants and examined 13 behavioral outcomes. The authors show that edges discarded by feature selection can still provide significant predictive accuracy and lead to different neurobiological interpretations. These findings hold across cognitive, developmental, and psychiatric phenotypes, and apply to both functional and structural connectomes.
Focusing solely on top features may present only the tip of the iceberg, while subtler, brain-wide signals—routinely ignored—can be equally meaningful. Accounting for these distributed signals can improve reproducibility, refine biomarkers, and better reflect individual variability in brain-behavior relationships.