Digital ‘Twin’ Brain Tech May Revolutionize Personalized Psychiatric Treatment

An AI model trained on individual brain scans predicted how people would behave on cognitive tasks with over 90% accuracy, then ran virtual experiments on the brain model to see how targeted changes might affect behavior.

Prescribing psychiatric medication has always involved a lot of guesswork. A drug that lifts one person’s depression may do nothing for another, and figuring out which is which usually means waiting weeks, sometimes months, for an answer. Researchers at Japan’s National Center of Neurology and Psychiatry think they’ve found a better starting point: a working digital copy of each patient’s brain that could one day help guide treatment decisions before anything is actually prescribed.

Published in BME Frontiers, the study describes an AI system that takes a brain connectivity scan and uses it to construct a personalized virtual brain. That model can then predict how a specific person performs on cognitive and emotional tasks and simulate what would happen if certain brain connections were altered.

At the center of the system is something called a connectome, a map of how different brain regions communicate with each other during rest, captured through standard brain imaging. Each person’s connectome is unique. Researchers feed that map into a two-part AI system: one component reads the connectivity data and uses it to build a personalized brain model for that individual, while the second simulates how the brain processes sensory input and produces behavior over time. Because the second component is built entirely from the individual’s own brain map, no two digital twins are the same.

Building and Testing the Digital Twin Brain

To develop and validate the system, the team drew on the Transdiagnostic Connectome Project, a publicly available dataset covering 228 participants, including 139 with psychiatric diagnoses spanning multiple conditions and 89 healthy controls, ranging in age from 18 to 68.

All participants had completed two tasks inside an MRI scanner. One, the Emotional Faces test, asked participants to match images of fearful or angry faces, measuring how the brain handles threatening emotional stimuli and how strongly the amygdala, the brain’s emotional alarm center, responds. A second task, the Stroop test, showed color words printed in mismatched ink colors (the word “blue” in red ink, for instance) and asked participants to name the ink color. Because the brain’s automatic reading response conflicts with the correct answer, it’s a well-established measure of cognitive control and mental processing speed.

Each participant’s data was split in half. The model trained on the first half, then was tested on the second, data it had never seen. On the Emotional Faces test, the digital twin matched participant behavior with 94% average accuracy. On the harder Stroop test, a three-choice task, accuracy held at 90%. Reaction time predictions and brain activity patterns also tracked closely with real measurements, and the model correctly showed heightened amygdala activity during emotional face processing, consistent with decades of neuroimaging research. Both accuracy and reaction time predictions outperformed prior models that attempted similar predictions from brain scans alone.

Running Experiments on a Virtual Brain

With accuracy established, researchers ran virtual experiments to see whether the model could simulate what might happen if specific brain connections were altered. No real brains were modified. Instead, they mathematically shifted each digital twin’s parameters toward two goals: reduced emotional reactivity and faster cognitive processing. For emotional reactivity, the simulated changes involved connectivity between limbic regions near the amygdala and parietal and temporal networks. For processing speed, the focus was on connections linking the prefrontal cortex, parietal lobe, motor cortex, and subcortical regions.

Most participants showed measurable shifts in the targeted function after the simulated change. Response sizes varied widely. Some showed large improvements, others minimal change, and a small number showed near-zero response. That spread closely mirrors the person-to-person variability clinicians see when real treatments are applied, which is what makes the model potentially valuable. Rather than predicting one uniform outcome, it reproduced the individual differences that make psychiatric treatment so difficult to plan.

What appeared to drive the variation was underlying biology rather than random noise. Participants whose baseline connectivity patterns aligned more closely with the direction of the simulated cognitive change tended to show stronger predicted responses, suggesting that a person’s brain wiring may shape how much they stand to benefit from a given intervention. It’s a promising lead, though the study stops well short of predicting outcomes for specific medications or therapies.

A Gap Between Simulation and the Clinic

A real-world gap remains between what the model can simulate and what current technology can physically do. Brain stimulation techniques like transcranial magnetic stimulation cannot yet target individual connections with the precision the model envisions. The researchers noted that targeting a small number of high-influence hub regions, areas whose activity ripples outward across broader networks, could produce effects closer to what the simulation predicts, an approach supported by recent work in stimulation-based network mapping.

The model also performed better at predicting behavior within the same individual than when applied to entirely new patients. Reaction time predictions for new participants dropped considerably, a limitation the authors attributed to the current dataset size. Larger datasets in future research may help address that.

For patients who have cycled through medications and therapies without understanding why results vary so much from person to person, a model grounded in individual brain biology is a meaningful step. Whether the digital twin brain eventually serves as a clinical planning aid, a drug research tool, or simply a new way to study psychiatric differences, it moves the field closer to treatment that starts with the patient’s own neurobiology rather than a population average.

Disclaimer: This study is based on computational simulations using data from 228 participants and has not been tested in clinical settings. The digital twin brain system has not been validated as a real-world treatment prediction tool. Findings should not be interpreted as evidence that this technology is ready for use in psychiatric diagnosis or treatment planning.