A NEW deep-learning framework using MRI scans offers promising accuracy in predicting brain age, cognitive decline, and amyloid plaque buildup, key factors in the early detection of Alzheimer’s disease and related neurodegenerative disorders.
Researchers developed a multi-modal artificial intelligence model trained on over 10,000 T1-weighted MRI scans from more than 7,000 individuals across six diverse cohorts. The model uses 3D convolutional neural networks to analyze imaging data and incorporates demographic variables through additional neural networks. This integrated approach allows for precise estimation of brain age and prediction of dementia-related changes.
The brain age model achieved a mean absolute error of just 3.302 years in cognitively normal participants from the ADNI dataset. Importantly, researchers observed that the discrepancy between predicted brain age and actual chronological age widened significantly as cognitive function declined, suggesting this model may serve as an early marker of neurodegeneration.
Extending this framework through transfer learning, the team created a cognition prediction model capable of assessing the Clinical Dementia Rating with a root mean square error of 0.334 and an AUC near 0.95 for identifying dementia. Notably, the model accurately identified affected brain regions, including the medial temporal lobe, which are critical in the progression of Alzheimer’s disease.
In the final arm of the study, the deep-learning model predicted amyloid plaque presence, a hallmark of Alzheimer’s, with an AUC of approximately 0.8 among dementia patients, further demonstrating the potential for MRI-based diagnostics in neurodegenerative disease management.
These findings underscore the potential of AI-driven MRI analysis as a non-invasive, scalable tool for screening and monitoring Alzheimer’s disease risk and progression in clinical practice.
Reference:
Wang C et al. Deep-learning based multi-modal models for brain age, cognition and amyloid pathology prediction. Alzheimers Res Ther. 2025;17(1):126.
