As hospitals and startups race to deploy AI-powered brain scans for diagnosing depression, predicting psychosis, and personalizing psychiatric treatment, a new study just dropped a very inconvenient truth: the algorithms powering these tools might be throwing away the most interesting parts of the brain.

Think of it like trying to understand an entire rainforest by only studying the tallest trees. Sure, you'll learn something. But you'll completely miss the fungi, the understory, the interconnected root networks doing the real heavy lifting beneath your feet. That, in essence, is what a team of Yale researchers found is happening across the field of brain-based machine learning.

Your Brain Is Not a Top-10 List

Here's the setup. When neuroscientists use machine learning to predict things like intelligence, ADHD symptoms, or psychiatric risk from brain scans, they're working with an absurd amount of data. A typical brain connectivity map - called a connectome - contains tens of thousands of connections between brain regions. That's a lot of wires to sort through.

So researchers do what any reasonable person drowning in data would do: they simplify. They use feature selection, a standard technique that keeps only the "most predictive" brain connections and tosses the rest. It's like a talent show where you only watch the top performers and assume everyone else was forgettable.

Brendan Adkinson, Dustin Scheinost, and colleagues decided to check on those rejected contestants. Using four massive datasets spanning over 12,000 brain scans and 13 different behavioral and clinical outcomes, they asked a beautifully simple question: what happens if you build models using the features that got cut?

The Understudies Stole the Show

The answer? Those "unimportant" brain connections could predict behavior just as well as the supposedly important ones. Let that sink in for a moment. The features that standard practice tells you to ignore were pulling their weight all along.

But here's where it gets really unsettling. The discarded features didn't just match the prediction accuracy of the top-ranked ones - they pointed to completely different brain networks. Different regions, different circuits, different neurobiological stories about what drives cognition, development, and mental health.

Imagine two doctors examining the same patient. One says the problem is in the heart, the other says it's the kidneys - and both are equally confident because both have equally strong evidence. That's essentially what's happening when different feature subsets tell contradictory neurobiological tales while performing identically as predictors.

This Isn't Just an Academic Headache

This matters beyond the ivory tower. The field of precision psychiatry is betting big on neuroimaging biomarkers to guide treatment decisions. If the brain networks we identify as "biomarkers" change dramatically depending on which statistical shortcut we take, we've got a foundations problem.

The findings held up across functional connectivity (measured by fMRI, which tracks blood flow as a proxy for neural activity) and structural connectivity (measured by diffusion tensor imaging, which maps the brain's white matter highways). They held up across cognitive measures, developmental outcomes, and psychiatric phenotypes. They held up in external validation datasets. This isn't a fluke in one corner of neuroscience - it's a pattern woven through the whole field.

Previous work has already sounded alarms about data leakage inflating prediction accuracy in connectome-based models, and about how brain-behavior predictions require enormous sample sizes to be reliable. This study adds another crack to the foundation: even when our predictions work, our interpretations of why they work might be telling us a story that's only half true.

The Tip of a Very Large Iceberg

The authors use a wonderfully apt metaphor: current practice reveals only the tip of the iceberg. The brain, that three-pound weather system between your ears, distributes its information processing across vast, subtle, overlapping networks. Squeezing all that complexity through a feature-selection bottleneck and then interpreting whatever survives as "the neurobiology" of a trait is like reading every third word of a novel and claiming you understood the plot.

The good news? Awareness is the first step. The study doesn't say feature selection is useless - it still helps build leaner, more practical models. But it does argue, convincingly, that we should stop treating selected features as the final word on which brain regions matter. The quiet, discarded connections carry signal too, and ignoring them may be fueling the reproducibility struggles that have dogged neuroimaging for years.

The brain, it turns out, doesn't organize itself for our statistical convenience. Every season of neuroscience seems to bring us back to the same humbling lesson: the more confidently we simplify the brain, the more it reminds us that its complexity is the whole point.

References:

1. Adkinson, B. D., Rosenblatt, M., Sun, H., Dadashkarimi, J., Tejavibulya, L., Horien, C., Westwater, M. L., Rodriguez, R. X., Noble, S., & Scheinost, D. (2026). Feature selection leads to divergent neurobiological interpretations of brain-based machine learning biomarkers. Nature Human Behaviour. DOI: 10.1038/s41562-026-02447-y

2. Rosenblatt, M., Tejavibulya, L., Jiang, R., Noble, S., & Scheinost, D. (2024). Data leakage inflates prediction performance in connectome-based machine learning models. Nature Communications, 15, 1829. DOI: 10.1038/s41467-024-46150-w

3. Marek, S., Tervo-Clemmens, B., Calabro, F. J., et al. (2022). Reproducible brain-wide association studies require thousands of individuals. Nature, 603, 654-660. DOI: 10.1038/s41586-022-04492-9

4. Shen, X., Finn, E. S., Scheinost, D., Rosenberg, M. D., Chun, M. M., Papademetris, X., & Constable, R. T. (2017). Using connectome-based predictive modeling to predict individual behavior from brain connectivity. Nature Protocols, 12, 506-518. DOI: 10.1038/nprot.2016.178

5. Botvinik-Nezer, R., Holzmeister, F., Camerer, C. F., et al. (2020). Variability in the analysis of a single neuroimaging dataset by many teams. Nature, 582, 84-88. DOI: 10.1038/s41586-020-2314-9

Disclaimer: The image accompanying this article is for illustrative purposes only and does not depict actual experimental results, data, or biological mechanisms.