AI Predicted Brain Amyloid From Blood Tests, but Faltered on New Patients

That is the main finding of a machine-learning study published in Frontiers in Neuroscience in 2026. It is a computational modelling study using existing data from people. It is not a clinical trial, and it did not test the model on patients in a doctor's office. The development data came from 299 participants: 170 from the 1Florida Alzheimer's Disease Research Center and 129 from the ADNI4 cohort, a national research study.

Amyloid PET is a brain scan that shows whether amyloid has built up. The authors note it is expensive, resource intensive and not available to everyone. That is why researchers are looking at blood tests as a cheaper, less invasive way to find out who is likely to have amyloid.

The researchers built a model that combined several kinds of information. These were blood biomarkers, cognitive test scores, APOE genotype (a gene linked to Alzheimer's risk), demographic details and measurements from structural MRI. The goal was to match the verdict of expert readers of the amyloid PET scan, labelled positive or negative.

They compared five types of classifier and several ways of choosing which features to use. The best performer was XGBoost, a tree-based method, paired with a featureselection step called mutual information. Under five-fold cross-validation, where the data are split so the model is tested on cases it has not seen in training, it reached a mean AUC of 0.891 (plus or minus 0.048) and a recall of 0.84. An AUC of 1.0 would be perfect and 0.5 is no better than a coin toss. Recall of 0.84 means the model caught 84 percent of the amyloid-positive cases.

The authors used a method called SHAP to see what drove the predictions. The strongest predictor was the ratio of plasma p-tau217 to amyloid-beta 42. Next came plasma p-tau217 itself, then p-tau181, the MMSE (a common memory and thinking screening test), and APOE ε4 status. P-tau217 and p-tau181 are forms of the tau protein measured in blood. Structural MRI measures added complementary information. In the authors' words, plasma tau and amyloid-related markers carried the strongest signal, while cognition, APOE and MRI refined the decisions.

The weak point came with outside testing. Evaluated on ADNI4, the model produced a mean AUC of 0.545, barely above chance. When it was trained on the Florida cohort and tested on ADNI4, the AUC was 0.636 and overall accuracy was 63 percent. The authors point to differences between the cohorts and a shift in the data between groups. They also report that p-tau217/Aβ42 and p-tau217 stayed the dominant predictors across cohorts, so the biological pattern held even though the predictions did not.

What does this mean for someone who is offered an AI-assisted Alzheimer's blood test result? First, the strong headline number came from the data the model was built on. Performance on new patients was much lower. A model that scores well in one clinic or research study may not carry over to another. Second, the study supports the idea that blood p-tau217 is a key signal. It does not show that an AI layer makes the result more reliable. Third, the study reports only amyloid PET visual read as the target. It does not show what happens to patients who receive such a result, and it is not a diagnosis tool.

The study does not offer guidance for patients, so what follows is our own suggestion, not the authors' advice. Anyone given such a result should discuss it with their clinician rather than act on it alone, and may wish to ask whether the test has been checked in more than one group of patients and whether the result would be confirmed with other testing.

The study is a reminder that brain health findings need care in reading. Our October 18 piece on ultra-processed food and silent brain damage on MRI covered a different study and a different question. Both show that a promising signal in one dataset is a starting point rather than a verdict.

The authors say future work may compare the model's predicted probabilities with Centiloid, a quantitative measure of amyloid load. The next thing to watch is whether the approach holds up when tested in several independent cohorts.

Removing Brain Immune Cells Restored Two Hours of Sleep in Alzheimer's Mice Gut Bacteria and Chemicals Differ in Mild Cognitive Impairment, Small Study Finds Starting Statins Soon After Diabetes Diagnosis Linked to Less Dementia Flavourings, Colourings and Sweeteners Linked to Higher Depression Risk Older Adults on Self-Chosen Vegan Diet Lost Leg Muscle in 12 Weeks Topical Vitamin K2 Cream Improved Wrinkles and Firmness in Small Trial Staying Active Over Years Linked to 44% Lower Frailty Risk East Germans Who Faced Covert State Repression Still Report More Stress Four Cardio Styles Compared: High-Intensity Intervals Beat Moderate Intervals on Fitness AI Predicted Brain Amyloid From Blood Tests, but Faltered on New Patients Boosting Gene ZNF865 With CRISPR Reversed Aging Signs in Degenerated Spinal Disc Cells Review of 26 Cohorts Links Time-Restricted Eating to Better Blood Sugar Years of Fitbit Data Link Steady Daily Rhythms to Slower Biological Aging Review of Nine Meta-Analyses Links Hearing Loss to 43% Higher Dementia Odds UK Biobank Study of 170,000 Women Links HRT to Higher Cardiovascular Risk Small Trial Finds Ultra-Processing Alone Did Not Change Short-Term Satiety Combining Cardio and Strength Training Beats Cardio Alone in Heart Rehab 20-Gram Protein Dose Failed to Boost Muscle Building in Older Men Mediterranean Diet Linked to Slightly Younger DNA-Based Aging Readings in Study Gut Microbes May Shape How Environmental Toxicants Affect the Brain, Review Says Review Says Simply Raising NAD+ or SIRT1 May Not Slow Ageing Review Calls Chronic Stress Before Surgery a Possible Modifiable Heart Risk Review Maps How CRISPR Delivery Could Target Aging's Hallmarks Deep Learning Model Melody Reads DNA Code to Predict Methylation Patterns Mostly-Easy Polarized Training Beat Threshold Workouts on Endurance Markers, Not VO2max Trump Executive Order Creates New NIH Division and Task Force on Women's Health Review of 46 Studies Finds Ultra-Processed Food Heart Research Methodologically Weak Review Finds No Consistent Epigenetic Age Reversal From Antioxidant Supplements Time-Restricted Eating Raises Ketones, Modestly Calms Inflammation, Review Finds Gut-Made Short-Chain Fatty Acids Linked to Brain Protection in Low Blood Flow