AI model reads brain connectivity to trace how networks reorganize with age - Bioengineer.org

Title: AI model reads brain connectivity to trace how networks reorganize with age

URL Source: https://news.google.com/rss/articles/CBMiogFBVV95cUxNYmxGQzZkR2kwbUx1eXI0eVV3dWN5cV9hcnpLVU9DdGtYbXRfM2s5Y3U4aDNGZGQyaHltOGdlUEFnWTExdTE5d3JTWXZnbVkxN3cxYjF4M0Nqbmx3a1F1b3hlVmZ0RWdpRnM0YXBJbi1qcms4ZjUwZ2lQR0M1azhqZWZWZ21WUFh6VXRWeW5leEZVMWg0WERvWXJPNHowekx5VUE?oc=5

Markdown Content: A new brain-age modeling framework links predictions of a person’s age from brain scans to specific functional connectivity circuits, revealing stage-dependent reorganization across large-scale networks during mid- to late adulthood. The study, published in Brain Informatics, was led by Yikang Cao, Marco Castellaro, Marco Zorzi and Alessandra Bertoldo.

To address this, the researchers developed a framework that integrates a novel neural architecture with an interpretability module. Resting-state fMRI data from 599 adults aged 36 to 95 years were used to compute whole-brain functional connectivity. The team built an Orthogonal Auxiliary Guidance Convolutional Autoencoder, or OAG-CAE, which learns age-relevant latent representations while separating out age-residual variation, followed by a regressor for brain-age prediction.

Brain-age models estimate how old a brain appears from its functional connectivity patterns, which are typically derived from resting-state functional MRI. These models have shown promise for characterizing large-scale age-related network changes and advancing understanding of neural aging mechanisms. However, existing approaches have faced a persistent challenge: achieving both robust predictive performance and stable identification of the connectivity signatures that drive predictions.

The key methodological innovation is a SHAP-guided PCA back-projection module. It applies SHAP, a technique for attributing a model’s output to its inputs, to PCA-reduced functional connectivity features, propagates the attribution through the latent representation and predictive pathway, and then projects the PCA-level attributions back into connectivity space. This allows the model’s age predictions to be localized to specific brain circuits rather than remaining opaque.

The resulting brain-age model achieved robust accuracy across mid- to late adulthood, with a mean absolute error of 6.308 plus or minus 0.566 years and an R-squared of 0.702 plus or minus 0.043. These figures indicate that the model captured a substantial share of age-related variance in functional connectivity while maintaining stable performance across the adult lifespan studied.