A general-purpose AI model built for reading spreadsheets can now help flag Alzheimer's disease from brain scans and lab results.
Researchers built M2PFN by grafting brain-scan and lab-data encoders onto TabPFN, a foundation model that was originally trained to make predictions from tabular data using in-context learning rather than fine-tuning. They trained those new encoders end-to-end while keeping TabPFN itself frozen, and added a contrastive alignment step that forces the two data types into a shared format the frozen model already understands. A gated shortcut also lets a simpler tabular-only prediction back up the combined result. On a 2,240-patient dataset from the Alzheimer's Disease Neuroimaging Initiative, the model sorted patients into cognitively normal, mildly impaired, or Alzheimer's categories with 65.55% macro-F1 and 82.21% macro-AUC, beating the other single-source and multi-source models it was tested against.
The bigger claim is generalization: the same architecture, with only its output head swapped, predicted cognitive test scores with a mean error of 1.743 points on a separate 1,250-patient group. It also held up on two outside hospital datasets, OASIS-3 and SCAN, without any retraining, even when the cognitive test used to score patients changed. Most medical AI models degrade sharply when moved to a new hospital, scanner, or patient population, which is the exact failure mode that has kept many diagnostic tools out of real clinical use.
This is still a research paper, not a cleared diagnostic device, and three test cohorts is a start rather than proof. But a model that keeps working when the hospital, the scanner, and even the test instrument change is a more useful signal than another leaderboard score.