July 7, 2026 | A new type of digital twin model of the brain, designed specifically for children with autism spectrum disorder (ASD), has been developed by researchers in Italy in hopes of enabling targeted treatment with transcranial magnetic stimulation (TMS) at the exact location in the brain where neural circuits are misfiring. The computational feat raises the possibility of reversing the neurodevelopmental condition and exemplifies an “AI for good” application to help counterbalance fears about nefarious uses of artificial intelligence, according to Lorenzo Gaetano Amato, Ph.D. student in BioRobotics at Scuola Superiore Sant'Anna in Pisa (Italy).
The digital twin has been termed a high-fidelity digital brain model (FEDE), the acronym of which means faith in Italian, he says. It serves as a platform for providing anatomically accurate brain models based on both functional data from electroencephalograms (EEGs) and structural data from magnetic resonance imaging (MRI) data.
In a head-to-head comparison with the standard digital twin approach, FEDE was found to do a better job of reconstructing the “frequency component” of the brain, what Amato describes as the notes composing the choral activity in the body’s most complex organ (PLOS Digital Health, DOI: 10.1371/journal.pdig.0001445). The technology made its debut by creating a brain digital twin of a toddler with ASD.
Neurons in the brain are separated by distance and coated with a certain amount of myelin, which has insulating properties affecting how quickly electrical impulses travel between them, Amato explains. This additional myelinization information is “usually absent from current computational models because it is not needed for the majority of brain conditions.”
But the information is regarded as critical with respect to ASD because distance alone only calculates basic pathways, not velocity and communication timing, he says. Without it, transmission delays can get drastically overestimated.
AI-powered digital twins are rapidly becoming an industry standard across most engineering disciplines and have most recently been used by defense contractors to create high-fidelity virtual replicas of military vehicles and aircraft, says Amato. They are now gaining momentum in the medical field in situations where it is not possible to use imaging to learn what is happening inside the body of a patient “in a particular context.”
In cardiology, for example, digital twins are used to map heart rhythms and plan surgeries and, in a research capacity, to test the effects of medications on the heart during physical activity. In the brain, digital twins are being highly researched for epilepsy care, he says, which is the application currently the most grounded in evidence.
In 20% to 30% of epilepsy cases where brain surgery has been performed, patients continue to experience seizures postoperatively. The mechanisms that originate epileptic seizures are well known— “In the end it’s just ... an overly excitable brain!”—but the remaining challenge where digital brain twins have been highly beneficial is locating sources (or foci) of seizure spreading.
Typically, an epileptic seizure begins from a precise point in the cortex and then expands to surrounding brain areas, but it is not easy to understand the starting point from a standard MRI. To gather details on seizure foci, intracortical recordings give the least noisy, most precise data via electrodes that are surgically implanted directly inside the skull and penetrate the brain tissue and are a unique source of data in the field of neurology, he says.
Though invasive, the intracortical data are superior to traditional EEGs for epilepsy because electrodes attached to the skull give rise to a trio of problems—volume conduction, which blurs and distorts signals; poor spatial resolution, making it hard to pinpoint exact sources; and signal attenuation, which hides deeper brain activity. Complicated engineering tactics are required to overcome those hurdles and understand where the signal is located, says Amato.
Digital twins can combine intracortical recordings with MRI to simulate different possible starting points and to see which one would allow a seizure to propagate. “This has been applied retrospectively to inform surgeons about which area of the brain to cut in patients suffering from drug-resistant epilepsy.”
With ASD, the idea is to instead blend the structural data from MRIs with the neural activity data from EEGs to fill the trial-and-error treatment gap, he says. Since the biological connection between those two aspects is unknown, FEDE can be used to simulate the various possibilities in terms of neurological parameters (e.g., synaptic efficiency and brain excitability) until it finds the combination that most closely recreates the empirical one observed on the actual EEG of a patient.
Even with Alzheimer’s disease, where the etiology is clearer relative to ASD, “it is hard to understand how the pathology progresses,” he adds. For this reason, relatively few computational models of the brain exist for either condition.
Unlike most of the people in his department, Amato doesn’t work on actual robots such as exoskeletons and robotic bioprostheses. A different approach is required when dealing with the brain to tailor a treatment to the peculiar state of the patient, particularly when it is impractical to use some sort of implant, he says.
Research on the use of deep brain stimulation for treating Parkinson’s disease, an approach involving surgical insertion of a conductive rod in the subthalamic nucleus, has been one of the main outputs of his lab over the past year. This cluster of neurons deep within the brain is deeply connected to the regions that control movements. The goal is clinical translation of the approach for tuning stimulation parameters, making them personalized to patient-specific targets.
The intention with FEDE was to create a computational model to improve understanding of ASD etiology as well as the “connection between structural aspects of the disease and its functional manifestations,” says Amato. Modeling the brain is nothing new but is generally done for conditions with pre-existing knowledge about how they start, progress, or can alter the structures of the major anatomical regions.
With ASD, the exact cause is murky since there are many co-occurring factors contributing to the landscape of the disease with no “well-defined evidence for bridging the different aspects,” he continues. Researchers therefore had to “go back to square one” to develop a pipeline for comprehensively analyzing MRI and EEG data with a single digital twin.
The FEDE approach involves the use of a “personalized conduction velocity map,” a 3D model that measures the speed at which electrical signals travel along specific white matter pathways in an individual's brain. Up to now, digital twins have modeled brain activity starting with the anatomical constraints of the brain mapped based on what an MRI reveals about where different structures are situated in the head of a patient, says Amato.
This gets back to one of the chief shortcomings of current methods for pinpointing problems in individuals with ASD, he says, since they don’t consider how different brain structures are communicating with each other. The time it takes a message to travel from one part of the brain to another, what’s called its conduction velocity, isn’t dependent solely on the distance between the two brain areas.
Factoring myelin into the equation addresses the problem. The insulating sheath that wraps around nerve fibers like the plastic coating on an electrical wire “makes up the external part of the cables of the axis that connect to neurons,” he explains.
To demonstrate the superior performance of FEDE over the standard digital twin approach, Amato and his team compared spectral properties and functional connectivity matrices obtained from their simulated EEGs. With the standard model researchers reported a loss of fidelity in reconstructing the functional aspect of the neuroactivity of the patient, as measured by a breakdown of brain wave signals into individual frequencies to show which ones were the strongest (“spectral density”).
“You can treat the brain activity of the patient as a sum of different sub-activities at different frequencies, to see the ratio between them,” Amato explains, drawing a parallel to decomposing the sounds of a guitar chord into the sum of its different notes. The standard approach had problems reconstructing both the frequency of these notes and the amplitude, meaning how much a frequency contributes to each note.
The dominant notes (frequencies) constituting the chord of a healthy brain are known, enabling a comparison of the frequency component in the brain of the toddler with ASD, he says. By that yardstick, the standard model significantly underestimated the contribution of the various notes.
While perfectly suited for ASD, FEDE could also be used for studying other types of conditions such as Alzheimer’s disease, Parkison’s disease, and multiple sclerosis where it is likewise essential to link structure and function. Development of a computational model for Alzheimer’s, which is the main topic of his Ph.D., “is different because ... we know to a certain degree what the structural alterations are and how they connect to the functional manifestations of the disease,” says Amato. The big problem here is understanding “when the structural degeneration will translate into an organ-specific morphology.”
How to translate the new digital brain twin into a tool for guiding TMS treatment of ASD is uncharted territory, or “terra incognita” as they say in Italy, Amato says. The next step is “to identify one particular feature of brain activity that can reliably differentiate autism spectrum disorder patients from typically developed children,” which would serve as the target for the distal therapeutic—much like the glycemic index is used by diabetes patients to guide food choices for keeping their blood sugar at a normal healthy value.
This ASD-specific index could be used in the future to reverse the condition with biosimulation, he adds. It should only be a matter of months before that target is found, since the research team has already started to analyze publicly available data from the Child Mind Institute containing several hundred MRI and EEG recordings from ASD children as well as a significant number of healthy control participants.
They are also collaborating with a nearby hospital in Tuscany for the ASD recordings, but the public repository will be invaluable in moving the project forward, he continues. It would otherwise take an inordinate amount of time to acquire a complete dataset. Parents also tend to be highly reluctant to allow their healthy children to undergo medical examinations for research purposes.
Longer-term plans include applying TMS retrospectively to a cohort of ASD children, each matched with a healthy child of a similar age and biological sex, to evaluate the impact of the treatment, says Amato.
Through computational modeling, researchers will then attempt to transform the EEG of the child with the autism spectrum disorder into that of its healthy twin.
The approach is intended to build a strategy for tailoring TMS treatment to the needs of ASD patients, a key objective of precision medicine. But FEDE, being a digital twin platform, could also logically have applications in the context of diagnostics and prognostic prediction, he notes.