New Deep Learning Tool Forecasts Alzheimer Disease Progression From Baseline MRI

Accurately predicting cognitive decline remains one of the most significant challenges in Alzheimer disease (AD) care and research. While advances in imaging, fluid biomarkers, and genetic testing have improved diagnostic confidence, forecasting disease progression often requires multiple assessments, specialized testing, and longitudinal follow-up that may not be readily accessible in many clinical settings.

Recently published in Nature Aging, a study from researchers at the University of California, San Francisco introduced a novel deep learning framework capable of predicting both current and future cognitive performance using only a baseline MRI scan and demographic information. Using data from the Alzheimer's Disease Neuroimaging Initiative (ADNI), the model demonstrated improved performance over existing AI approaches while eliminating the need for PET imaging, fluid biomarkers, genetic testing, or extensive neuropsychological assessments. Investigators reported that the approach successfully predicted clinically relevant outcomes, including cognitive scores, AD diagnosis, and future disease trajectory from a single imaging study.

In an interview with NeurologyLive^®, senior study author Ashish Raj, PhD, professor of radiology and biomedical imaging at UCSF, discussed how clinicians might ultimately incorporate the tool into routine practice, its current limitations, and its potential applications in clinical trial enrichment. He also addressed important considerations regarding real-world implementation and ongoing efforts to validate the model using community-based MRI datasets.

NeurologyLive: How would a neurologist practically use this tool today? At what point in the diagnostic workup would a baseline MRI-based prediction realistically slot in?

Ashish Raj, PhD: Since our tool uses only MRI, most suspected dementia patients will either already have one or can be easily and cheaply prescribed one. Then the tool will predict their cognitive score at the present and 2-3 years in the future. This will give the neurologist sufficient information to determine the course of treatment or additional testing. Hence, the most likely stage at which this tool will be useful is in the earliest or suspected cases of dementia.

What are the biggest limitations clinicians should know about before trusting this model's predictions? Are there patient populations where it performs less reliably?

Yes, absolutely. This tool is only validated on Alzheimer's spectrum subjects. Hence, it should not be used to rule out other diagnoses or assess cognitive impairment in non-Alzheimer's dementias.

You mention potential use in clinical trial design for disease-modifying drugs. What would a trial using this tool actually look like compared to current patient selection methods?

Most trials fail because there is not sufficient effect size, which relates to sample size and therefore to cost and complexity of the trial. This tool can be used at intake time to assess whether the subject has or is likely to have cognitive impairment without the drug intervention. Then only those subjects should be recruited in the trial or order to allow higher effect sizes due to the drug in the most cognitively vulnerable subjects.

What would need to happen, regulatory or technical, before a community neurologist could run this on their MRI scanner?

This tool is an assisted reporting system which is not a diagnostic tool in itself. Hence, it likely will not be regulated heavily. A neurologist is free to use it for their decision making at will. Technically, there is no need for any specialized tool or hardware at all. This tool can run either on a personal computer or on the cloud, depending on the business model.

The model was trained on ADNI data, which skews toward research-grade imaging. How well do you expect it to hold up on variable-quality scans from community hospital settings?

This is a good question to which we currently do not have an adequate answer. We believe that most clinical standard of care T1 MRI scans are of sufficient quality to allow this tool. But only time and additional testing can actually answer this. We are working on exactly this angle at UCSF, using our clinical scan databases to study this aspect.