AI researchers are investigating whether patterns in medical records, images and other measurements can help identify neurodegenerative disease earlier. The promise is worth following. It does not mean a chatbot or an everyday app can tell you whether you will develop dementia.
A research model predicts an outcome from data. A diagnosis involves evaluating a particular person. Confusing those tasks can create unnecessary fear or false reassurance, especially when a headline makes a precise prediction sound inevitable.
What the “seven years earlier” study did
A 2024 UCSF study published in Nature Aging used electronic health records to develop models predicting Alzheimer’s disease onset. Its analysis included records from 749 people with Alzheimer’s and 250,545 controls, examining information available at different times before the study’s defined onset point.
The paper reported a mean AUROC of 0.72 at seven years beforehand for one model. AUROC describes how well a model ranks cases relative to controls across thresholds. It does not mean that 72% of people given a warning will develop Alzheimer’s, or that everyone can be diagnosed seven years before symptoms.
The researchers also investigated associations involving conditions such as high cholesterol and osteoporosis. An association can suggest a question for further research. It does not establish that one condition causes another, or that treating a predictor prevents the predicted disease.
Prediction, screening and diagnosis differ
A prediction estimates a possibility. Screening looks for indications that further evaluation may be appropriate. A diagnostic assessment considers what explains someone’s current symptoms. Each use needs evidence suited to its purpose.
A tool trained to distinguish two carefully selected research groups may behave differently among people in an ordinary clinic. Language, recording quality, missing records and differences in the people tested can matter. Ask whether the model has been evaluated in the setting where someone proposes to use it.
Most importantly, ask what happens after a result. A useful clinical tool needs a clear next step and evidence that using it helps care. An unexplained risk score with no appropriate follow-up can leave a person worried without giving them useful guidance.
Even an authorized test has boundaries
AI prediction is also different from measuring a biological marker. As one example, the FDA’s May 2025 clearance announcement for the Lumipulse G blood test described an aid for evaluating amyloid pathology in symptomatic adults aged 55 and older. The announcement explicitly said the test was not intended for screening or as a stand-alone diagnosis.
That example is not a recommendation that every reader obtain that test, nor a complete list of available tests. It illustrates why the exact intended use matters. A clinician interprets a result with other information, and the right evaluation depends on the person.
Read accuracy claims with the denominator
Consider a fictional example, unrelated to a particular product. A study reports that a tool identifies most affected participants correctly. To understand it, you also need to know how often unaffected participants receive an incorrect warning, how many people were tested and how common the condition was in that group.
Ask these questions before trusting a new early-detection claim:
- Was the outcome a confirmed diagnosis, a future recorded diagnosis or a different marker?
- Did the researchers test on people whose data were not used for training?
- Were results checked in another clinic or population?
- How were false alarms, missed cases and uncertain results reported?
- Does the tool have authorization for the use being advertised?
A broad claim that “AI detects dementia” hides those distinctions. Results for Alzheimer’s should not automatically be extended to Parkinson’s disease or every other neurological condition.
What to do with a personal concern
If changes in memory or thinking concern you, arrange a medical evaluation rather than relying on a chatbot prediction. The National Institute on Aging recommends discussing concerning changes with a doctor.
Prepare concrete examples: what changed, when it began and how it affects everyday activities. An AI tool may help organize a nonsensitive list of questions, but avoid uploading another person’s health records without appropriate permission and privacy protections.
Research into earlier detection is valuable precisely because the questions are difficult. Follow the evidence without turning a statistical association into a diagnosis or a research milestone into a personal forecast.
Updated September 23, 2026. This AI-assisted article explains the linked sources and includes clearly labeled illustrative examples. It is for general information, not medical advice, diagnosis or treatment. See our disclaimer.
Featured photograph: Robina Weermeijer via Unsplash. Anatomical teaching model; not a patient scan or diagnostic result.
