Alzheimer's AI model SCSR hits 0.848 AUC across Chinese and UK populations

Stochastic cortical self-reconstruction (SCSR) is a new AI approach for personalized mapping of gray matter atrophy, a key biomarker for Alzheimer's disease (AD). Unlike conventional normative modeling that operates on coarse brain regions and is limited by training covariates, SCSR estimates an individualized healthy reference directly from cortical thickness at the vertex level. This enables detection of subtle, subject-specific deviations from healthy cortical shape. The method was originally trained on UK Biobank (UKB) data, and this paper from researchers including Fabian Bongratz and colleagues tests how well it generalizes to an independent Chinese population.

Comparing four training strategies—direct application, fine-tuning on Chinese data, training from scratch, and joint training—the researchers evaluated discriminative performance across healthy, MCI, and AD groups. Using both multilayer perceptron (MLP) and Spherical UNet (SUNet) backbones, the fine-tuned SUNet achieved the highest performance with an average pairwise AUC of 0.848, closely followed by the UKB-trained SUNet. Crucially, reconstruction errors remained low across the entire lifespan, even when the training population had a much narrower age distribution. This demonstrates that SCSR not only transfers across ethnic populations but also remains robust across aging, making it a strong candidate for clinical deployment in diverse global settings.