The Crumbling Connectome: How Engineers Built a Dementia Workbench in Silicon

A Scene from Two Worlds at Once
A patient sits in a clinic, having a difficult time recalling the name of a close friend. Across the world, in a laboratory in Chengdu, China, a tiny integrated circuit quietly hums. Its hundreds of electronic nodes lose contact with one another in a sequence that echoes what is happening inside that patient’s skull. In doing so, the chip captures patterns of disconnection that parallel those seen in degenerative brain disease. No neurons are involved, no biological tissue, no living cells. Yet researchers designed Fushighi-1, a CMOS device, so its network can fall apart in a controlled way. That lets them observe how coordination among regions fails as connections disappear, step by step, in Alzheimer’s disease and Frontotemporal dementia.
A group from institutions across China, Japan, Italy, Cameroon, and Cuba has now published a framework that maps neurodegenerative disease progression onto a physical electronic circuit. That circuit models the actual wiring of the human cerebral cortex. The work, accepted in IEEE Access, merges neuroscience, circuit engineering, and clinical insight into one reproducible hardware platform.
The Brain as a Network
To understand what this chip does, think of the brain not as a single organ but as a vast web of interconnected regions. Neuroscientists call this the connectome, essentially the brain’s wiring diagram. It shows which regions communicate with one another, mapping the anatomical pathways that link every area of the cortex. When those pathways stay healthy and intact, brain regions coordinate fluidly, supporting memory, movement, attention, and emotion. When they degrade, as happens in neurodegeneration, communication frays, and cognition deteriorates in step.
Alzheimer’s disease typically erodes memory and spatial orientation first. Frontotemporal dementia, by contrast, often reshapes personality, behavior, and language well before memory loss becomes obvious. That clinical distinction matters because each disease follows a different spatial route through the brain. Alzheimer’s involves the abnormal accumulation of amyloid plaques and misfolded tau proteins that progressively damage circuits connecting memory and cognition centers. Frontotemporal dementia spreads toxic protein deposits, particularly TDP-43 and tau, through the frontal and temporal lobes. Both conditions follow recognizable trajectories through the connectome, and researchers have spent decades charting those trajectories using MRI, tau-sensitive PET scanning, and post-mortem tissue analysis. The Chengdu group asked whether that map could take circuit form, letting scientists study downstream effects on brain coordination without waiting for clinical trials or animal experiments.
How the Circuit Works
The chip uses ring oscillators rather than neurons. These are loops of circuitry that naturally generate rhythmic electrical signals, serving as a stand-in for the activity that real brain regions produce. Each of the 360 nodes on the chip corresponds to a distinct cortical region, as the Human Connectome Project’s multimodal parcellation atlas defines them, a research-grade anatomical map built from thousands of brain scans.
The chip’s architecture reflects that atlas directly. Connections between nodes follow structural connectivity data from diffusion tractography, a neuroimaging method that traces fiber bundles running through white matter. The result is a sparse but organized network with a density of about five percent. As in the real brain, most regions do not connect directly to each other but remain reachable through a few intermediate steps.
Each oscillator node generates a wide range of dynamical behaviors depending on coupling strength and applied voltages. Under the right settings, the nodes produce chaotic signals, complex yet stable, sitting between rigid periodic rhythms and unstructured noise. That controlled complexity is deliberate. The brain operates in a zone of rich, nonlinear dynamics, and the chaotic oscillators on Fushighi-1 occupy a comparable regime.
Coupling between nodes is controlled electrically. Adjusting a node’s selection voltage severs its connections, leaving it to oscillate locally but unable to influence or respond to its neighbors. This resembles a brain region that stays active but loses its long-range anatomical connections. That isolation is the circuit’s analog of neurodegeneration.
Building the Disease Timelines
To orient both audiences from the outset the researchers built several disease timelines from different data sources Inclusive protein imaging, structural brain scans, and post-mortem tissue examination, because different tools reveal different aspects of the same underlying process. These timelines took the form of disease progression vectors, lists indicating which of the 360 nodes the disease had reached at each step. The researchers assembled five distinct damage sequences for Alzheimer’s disease and two for behavioral variant Frontotemporal dementia.
Three sequences covered Alzheimer’s. The first drew from connectome analyses of tau pathology, tracing how tau accumulation disrupts inter-regional communication. The second used tau-PET imaging data showing retention patterns across the aging brain, calibrated against established Braak staging of neurofibrillary tau pathology. The third derived its staging from longitudinal MRI databases tracking cerebral atrophy across the lifespan, with stages defined by statistical groupings from imaging data rather than clinical milestones alone.
For Frontotemporal dementia, one sequence followed immunohistochemical observations of TDP-43 inclusion patterns from post-mortem autopsy cases. The other used the same MRI-based atrophy framework applied to FTD cohorts.
Each stage in each sequence translated into a specific chip configuration, with certain nodes isolated and others left active, coupled, and free to interact. The team then ran detailed circuit simulations using a professional SPICE environment, recording voltage signals from all 360 nodes at every configuration. Computing initial conditions from operating-point analysis kept results consistent across runs.
What the Measurements Revealed
From those recordings, the researchers computed three families of metrics. The first was phase synchronization, a measure of how often pairs of nodes keep time together. It served as a circuit-level surrogate for the functional connectivity that fMRI studies measure. The second was permutation entropy, which reflects how rich and varied each node’s signal is, acting as a proxy for the complexity of local dynamics. The third was dominant frequency, tracking whether activity across the network leaned toward faster or slower rhythms.
The same broad trends appeared across every disease condition and every damage sequence. As the team severed structural connections progressively, average synchronization across the network fell. Isolated nodes showed very low phase coherence with the rest of the network. Those still connected experienced a gradual decline in their coordination with one another.
Signal complexity also fell as damage accumulated. This mirrors a phenomenon long recognized in EEG studies of dementia patients, where brain electrical activity becomes less varied and less rich as disease advances. The spectral response shifted in tandem. Under healthy conditions, all 360 nodes operated near the same frequency. As nodes dropped out, the remaining active nodes drifted toward lower average frequencies, loosely analogous to the EEG slowing seen in early Alzheimer’s disease, where the dominant rhythm shifts from the faster alpha band toward the slower theta range.
Reading the Graph Metrics
Graph-theoretic analyses of the simulated functional connectivity added further depth. Global efficiency, a measure of how quickly information can travel across the entire network, declined steadily as structural damage mounted. The clustering coefficient, indicating how tightly nodes form local groups, fell as well. Modularity, an index of how clustered the network becomes into separate, internally synchronized communities, rose in the late stages of several damage sequences. This pattern is consistent with clinical observations of abnormal network fragmentation in both diseases. Together, the findings point to a loss of balance between local specialization and global coordination, a defining feature of diseased brain networks.
Disease-Specific Signatures
Each disease timeline reshaped the network in a distinct way, and that distinction reveals how much the starting point and path through the connectome determine what fails first. This variability was one of the most informative aspects of the results.
In the Alzheimer’s sequence informed by tau connectome data, the Default Mode Network and Fronto-Parietal Network showed early vulnerability. This aligns with the well-documented involvement of the posterior cingulate cortex and prefrontal regions in Alzheimer’s pathology. The visual and somatomotor networks held up considerably longer, consistent with clinical observations that primary sensory areas are generally spared until late stages.
The tau-PET sequence produced a different picture. Early disruptions concentrated in the limbic and visual networks before a sudden collapse of the Default Mode Network at a middle stage. Some networks even showed a transient increase in synchronization before ultimately declining. Some regions briefly became more coordinated than usual, as the circuit appeared to sustain overall performance despite mounting damage, an effect the researchers link to compensatory hyperconnectivity reported in early Alzheimer’s patients. The eventual collapse of the Fronto-Parietal Network in later stages may mark the point at which that compensatory capacity runs out.
In both Frontotemporal dementia sequences, the limbic network and Fronto-Parietal Network deteriorated earliest. Visual regions remained relatively intact across both FTD conditions, consistent with the clinical reality that FTD patients rarely present with visual symptoms. Between the two FTD sequences, the TDP-43-informed timeline produced faster and more catastrophic disconnection, while the MRI atrophy-based sequence followed a slower, more gradual decline, a contrast reflecting the differing natures of the underlying pathological processes each one models.
These differences across sequences are a feature, not a weakness. They make explicit that the measurement tool, whether tau deposition, atrophy volume, or functional connectivity change, shapes the disease trajectory one recovers. All of this emerges on the same hardware platform under controlled conditions.
What This Means for Medicine and Engineering
The implications reach in several directions at once.
For researchers studying neurodegeneration, Fushighi-1 offers something software-based models do not always provide. Other labs can run the same chip or circuit description independently and obtain the same behavior. That reproducibility matters greatly in a field where experiments are often difficult to replicate across groups. Because the research group has made the chip’s design files and SPICE netlists publicly available, any group with access to the same simulation tools can reproduce the full set of results exactly.
For biomedical engineers, the platform opens a path toward hardware-in-the-loop testing, an approach where algorithms run against signals from actual circuits rather than purely computed data. This can expose failure modes and sensitivities that software alone might miss. The electronic biomarkers the circuit generates, including changes in synchronization, entropy, and spectral composition, have direct analogs in the clinical EEG and fMRI literature. That creates a conceptual bridge between what the chip reports and what clinicians currently measure in patients.
For those working on neuroprosthetics and brain-computer interfaces, the chip may serve as a testing environment for closed-loop systems. These devices sense ongoing brain activity and adjust stimulation in real time rather than following a fixed pattern, allowing treatment to adapt continuously to a patient’s current neural state. Having a physical circuit that produces brain-like dynamics across a range of disease conditions, without requiring animal experiments or human subjects, could substantially accelerate the development and validation of such systems.
What the Model Cannot Do
It is important to be clear from the outset that the researchers model coordination patterns, not the cellular mechanisms of disease. The oscillators are not neurons. They do not fire action potentials, release neurotransmitters, or encode information the way biological circuits do. The functional connectivity the chip produces is a synchronization-based approximation of fMRI-derived connectivity, not a precise replica.
Some patterns the chip generated diverged from typical clinical presentations. In one Alzheimer’s sequence, the visual network showed early decline, which is more characteristic of the posterior cortical atrophy variant than of typical Alzheimer’s disease. In another, the limbic network appeared more resilient than clinical experience would suggest, since memory-related limbic structures are classically among the first affected. These discrepancies reflect the difficulty of capturing all aspects of a multifactorial disease with a single choice of tuning values for the oscillators and couplings. The authors acknowledge that adjusting those parameters for each modality individually remains an area for future development. Improvements in how the team constructs the disease progression vectors could further enhance the biological plausibility of the observed large-scale behavior.
A Physical Workbench for the Brain
The contrast with existing tools is worth making explicit. Spiking neuromorphic architectures such as TrueNorth and SpiNNaker focus on individual neuron-like events and microscale dynamics. Fushighi-1 concentrates on whole-region dynamics and large-scale network behavior. Software platforms such as The Virtual Brain perform mean-field simulations of connectome-level activity on general-purpose computers. Fushighi-1 occupies a different niche, a macroscale, analog, physically instantiated model of the whole-brain connectome that researchers can reconfigure to simulate any neurological condition for which a spatial staging progression exists.
The ability to reconfigure which nodes are active or isolated is what makes this approach particularly powerful. It gives researchers a way to test many different disease maps on the same piece of hardware. The 360 nodes are not abstractions in a computer’s memory. They are real transistors, real oscillators, physically connected according to real human brain anatomy. When a node is isolated, the electronic equivalent of a long-range anatomical connection is cut, and the dynamics of the remaining network shift in ways that echo what happens inside a degenerating human brain.
By studying how a chip forgets, researchers may better understand how to detect early signatures of dementia or evaluate strategies to slow its spread, even if direct treatment ultimately depends on biology. The fact that a semiconductor circuit can capture something true about the slow catastrophe of neurodegeneration represents an unexpected convergence. The brain and the transistor are both information-processing systems shaped by physics, and in certain regimes they speak something like the same dynamical language. This research, built on that shared vocabulary, suggests that silicon may yet have something to teach us about forgetting.
Citation
Paul Didier Kamdem Kuate, Joakim Vianney Ngamsa Tegnitsap, Zeric Tabekoueng Njitacke, Manyu Zhao, Theophile Fonzin Fozin, Pedro A. Valdes-Sosa, Hiroyuki Ito, Mattia Frasca, Federico D’Agata, and Ludovico Minati. “Electronic emulation of neurodegenerative disease progression on a CMOS-based physical model of human connectome.” IEEE Access, DOI: 10.1109/ACCESS.2026.3670097. Accepted 2026.
