Overlooked Brain Connections Hold Clues to Cognition and Mental Health

Amazing stuff!

"Key points

• Scientists who use imaging to understand the brain’s complexity often focus on the strongest signals and discard the rest.
• A new study reveals that connections routinely overlooked as “noise” during neuroimaging data analysis can predict behavior with remarkable accuracy.
• The finding could help explain why some people with psychiatric illness don’t respond to treatments, and it could identify new targets for therapeutics.

...

For the study, researchers investigated whether signals discarded by feature selection could reveal meaningful insights about brain and behavior. The team examined brain imaging and behavioral data from more than 12,000 participants across four major U.S. datasets. For every participant, the team calculated the strength of association between brain connections and the outcome they wanted to predict.

All the connections were then ranked from the strongest to weakest associated and divided into 10 non-overlapping groups.

Group one contained the top 10% of connections, those that scientists usually select, while groups two through 10 held the remaining 90% of connections—the connections often dismissed as noise. The team then built 10 prediction models, one for each group. ...

The team found that lower-ranked connections—groups two through nine—consistently achieved prediction accuracy similar to the top 10% of connections. 

In some cases, models built on lower groups of connections performed better than those trained on the top group. The authors suggest this might be because predictive information is widely distributed throughout brain connections and not just concentrated within the strongest ones. ..."

From the abstract:

"A central objective in human neuroimaging is to understand the neurobiology underlying cognition and mental health.

Machine learning models trained on neuroimaging data are increasingly used as tools for predicting behavioural phenotypes, enhancing precision medicine and improving generalizability compared with traditional MRI studies.

However, the high dimensionality of brain connectivity data makes model interpretation challenging. Prevailing practices rely on selecting features and, implicitly, interpreting identified feature networks as uniquely representative of a given phenotype while overlooking others. Despite its widespread use, how univariate feature selection balances the trade-off between simplification for optimizing modelling and oversimplification that misrepresents true neurobiology remains understudied.

Here, using four large-scale neuroimaging datasets spanning over 12,000 participants and 13 outcomes, we demonstrate that edges discarded by feature selection can achieve significant prediction accuracies while yielding different neurobiological interpretations. These results are observed across cognitive, developmental and psychiatric phenotypes, extend to both functional connectivity (functional MRI) and structural connectomes (diffusion tensor imaging) , and remain evident in external validation.

They suggest that focusing on only the top features may simplify the neurobiological bases of brain–behaviour associations. Such interpretations present only the tip of the iceberg when certain disregarded features may be just as meaningful, potentially contributing to ongoing issues surrounding reproducibility within the field. More broadly, our results reinforce that subtle brain-wide signals should not be ignored."

Overlooked Brain Connections Hold Clues to Cognition and Mental Health | Yale School of Medicine

Feature selection leads to divergent neurobiological interpretations of brain-based machine learning biomarkers (open access)

Fig. 1: CPM [connectome-based predictive modelling ] across non-overlapping decile-ranked brain connectivity features.

a, Workflow illustrating the decile-based CPM pipeline, including the initial correlation of connectivity features with phenotypic outcome, ranking features on the basis of group-level correlations between edges and phenotype, splitting features into deciles, and evaluating each decile-based model. DTI, diffusion tensor imaging.

b, Violin plot showing the predictive performance of models trained on each decile of features within the PNC dataset for executive function.

c–e, Radar plots depicting predictive performance across deciles for PNC executive function (c, left), PNC language abilities (c, right), HCPD executive function (d, left), HCPD language abilities (d, right), HBN executive function (e, left) and HBN language abilities (e, right). Bold decile numbers indicate significant predictions.