Cohort-Amortized Personalization Builds Brain Twins 100x Faster, Privately

Personalized generative brain models require individual neuroimaging data, but privacy constraints and re-identification risks make sharing difficult, while per-subject fitting takes hours of compute – limiting clinical translation and multi-site studies. To solve this, researchers introduced Cohort-Amortized Personalization (CAP), a method that replaces data sharing with model sharing. A neural density estimator is trained on simulations from a mechanistic whole-brain model under a low-rank cohort prior. Only the compact estimator is distributed, so new subjects are personalized in seconds on their own data alone. CAP also uses a Cross-Autoencoder (CrossCoder) to map connectomes from 20 different anatomical atlases into a shared latent space, making the method atlas-independent and deployable across sites with heterogeneous protocols.

The researchers validated CAP on two cohorts: 21 patients with drug-resistant epilepsy (epileptogenic-zone localization F1=0.56) and 832 subjects from the 1000BRAINS aging cohort (predicted age correlation r=0.44). In both cases, CAP matched or exceeded per-subject inference while achieving hours-to-seconds speedups. The shared artifact couples a cohort prior to a mechanistic simulator, enabling in-silico experimentation and synthetic-cohort generation without raw-data access – a governance-audited alternative termed 'synthetic access.' This approach paves the way for wider adoption of personalized brain modeling in diverse clinical settings, from epilepsy surgery planning to aging research.