AI finds Alzheimer's brain networks need personalized stimulation, not single hotspots

Researchers led by Cristiano Capone at the Italian National Institute of Health (ISS) developed subject-specific reservoir-computing models that reconstruct each individual's lagged functional connectivity (FC) from resting-state fMRI. They tested whether Alzheimer's disease (AD) FC signature reduces to a few focal sites—a key question for targeted neuromodulation like deep brain stimulation. Their models, trained to distinguish AD from healthy controls, achieved modest accuracy (below structural atrophy) but crucially allowed in-silico simulation of stimulation effects.

The team found that the ideal correction to map a patient's dynamics onto a control template is a distributed change in the model's connectivity kernel—a coordinated multi-site pattern. Stimulating the single node with the largest kernel deviation failed to revert classification, even at supra-physiological amplitudes. However, when they selected each patient's stimulation site based on its “effect on the disease discriminant” (i.e., how much it moves the classification score toward healthy), a single-site closed-loop controller achieved complete, individualized reclassification at lower dose. Optimal targets were cortical and highly heterogeneous, showing that effective neuromodulation requires model-informed, personalized targeting: the best site is where the network is most therapeutically responsive, not where the read-out deviation is largest.