A new AI approach may significantly simplify the prediction of Alzheimer’s disease progression. Researchers from the University of California San Francisco have developed a deep learning model that can estimate cognitive decline using only a baseline MRI scan combined with basic demographic data, thus eliminating the need for extensive cognitive testing.
Alzheimer’s disease accounts for an estimated 60–70% of dementia cases worldwide, affecting millions each year. Yet predicting disease progression remains complex, typically requiring a combination of imaging, biomarkers and time-intensive neuropsychological assessments. Limited access to such testing further complicates early diagnosis, particularly outside specialised centres.
Multitask AI model
The UCSF team introduced a multitask deep learning framework that integrates domain-specific knowledge into the model architecture. A key innovation is the inclusion of a brain tissue segmentation task, classifying MRI images into grey matter, white matter and cerebrospinal fluid, alongside prediction tasks.
By learning these related tasks simultaneously, the model builds a richer understanding of brain structure. This approach addresses a longstanding limitation of standard AI models, which often struggle to capture the variability of Alzheimer’s progression when relying on MRI data alone.
In the study the model outperformed existing AI methods, including transfer learning approaches commonly used with limited datasets. It accurately predicted not only Alzheimer’s diagnosis, but also current and future cognitive scores, all from a single MRI scan.
Eliminating reliance
According to senior author Ashish Raj, the model’s main advantage lies in its accessibility. “Unlike previous approaches, our model does not require baseline cognitive assessments, PET scans, genetic analysis or fluid biomarkers,” he noted. “This makes it faster, more cost-effective and easier to implement in routine clinical settings.”
Current neurocognitive testing can be both time-consuming and resource-intensive, often requiring trained specialists. By contrast, MRI is widely available in clinical practice. Leveraging this existing infrastructure could lower barriers to early detection and monitoring.
First author Daren Ma added that the method also reduces reliance on specialised MRI analysis software. This could accelerate clinical workflows and support earlier decision-making, particularly in settings with limited expertise.
Validation across datasets
To train and validate the model, researchers used data from the Alzheimer’s Disease Neuroimaging Initiative, supplemented with scans from the Human Connectome Project and the Dallas Lifespan Brain Study. This combination of datasets improved the model’s robustness and generalisability.
Importantly, exposure to brain scans from healthy adults helped the model distinguish between normal ageing and disease-related changes. The result is a system less prone to segmentation errors and better suited for real-world application.
The model also demonstrated the ability to forecast longitudinal cognitive decline based on baseline data alone. This is an advance that could streamline both clinical care and research.
Neurodegenerative care
Beyond Alzheimer’s disease, the researchers believe the approach could be extended to other neurodegenerative conditions, including Parkinson’s disease, ALS and Huntington’s disease. Predicting cognitive decline from minimal input data could be particularly valuable in community settings where access to specialised testing is limited.
The ability to identify patients likely to experience rapid progression may also benefit clinical trials. “Predicting progressors from baseline data could reduce sample sizes and costs in trials of disease-modifying therapies,” Raj explained. Future iterations of the model may incorporate additional data sources, such as longitudinal imaging, PET scans or biomarkers, to further enhance predictive accuracy. However, the researchers emphasise that real-world clinical validation remains essential.
If successfully implemented, this AI-driven approach could mark a significant step toward more accessible, scalable and timely diagnosis of Alzheimer’s disease, helping clinicians intervene earlier in the disease trajectory.
UF-MRI
Recently, researchers at the Champalimaud Foundation developed a new ultrafast MRI technique that reveals, for the first time, the direction of information flow in the brain. Using UF-MRI and the uFLARE method, they achieved millisecond-level temporal resolution and sub-millimetre spatial precision in rat brains.
The study showed how signals travel either from sensory input to higher brain regions (feedforward) or in the opposite direction (feedback). By combining high-resolution imaging with computational models, researchers mapped these signal pathways in the layered structure of the visual cortex.
The findings provided a new perspective on brain function, focusing not just on where activity occurs, but how information flows, and may improve understanding of neurological and psychiatric disorders such as Alzheimer’s, autism and schizophrenia.
