arXiv
An Alzheimer’s model adapts without retraining
An Alzheimer’s model can adapt to a new hospital by changing its reference patients instead of retraining its weights. Researchers trained M2PFN on 2,240 ADNI subjects and tested it with reference sets from OASIS-3 and SCAN without updating the model. If those local cases are representative and the system is clinically validated, a memory-clinic radiologist compares a patient with approved patients from the same setting.
An Alzheimer’s prediction model does not automatically travel well. A system trained on one research cohort may struggle when another hospital uses different scanners, sees a different mix of patients, or records different clinical information. Today, the usual answer is to retrain or heavily fine-tune the model. That is costly, and it can leave a clinic with a score whose accuracy does not travel with the patient.
Why one hospital’s model misses another hospital
A structural MRI scan can show changes in brain volume linked with neurodegeneration. But a scan is only part of the picture. Age, genetic risk, medical history, and fluid biomarkers can add information that an image alone cannot provide.
The problem is that hospitals do not collect these clues in exactly the same way. One dataset may contain many clinical variables, while another has only a few. Scanner hardware, image-taking procedures, diagnostic habits, and patient populations can all differ.
The new system, called M2PFN, combines a three-dimensional MRI scan with this table of clinical information. It can sort people into three groups—cognitively normal, mild cognitive impairment, or Alzheimer’s disease—or estimate a cognitive test score.
Its unusual ingredient is an in-context model. Instead of relying only on knowledge fixed during training, this kind of model compares a new patient with labeled examples supplied when it makes a prediction. The examples are a reference set: patients whose diagnoses or cognitive assessments are already known.
That reference set is not a magic folder. It requires trustworthy local labels, collected and standardized for reuse.
The model learns to use new local examples
The researchers trained M2PFN on 2,240 subjects from the ADNI research cohort. The model’s central tabular prediction engine, called TabPFN, stayed frozen. The parts that turn MRI scans and clinical tables into useful features were allowed to learn.
The key step was making the path into the frozen engine differentiable. In plain language, the researchers changed the input route so that the system could send feedback backward through it during training. MRI and clinical-data encoders could then learn to produce features that TabPFN could use, even though TabPFN itself was not being rewritten.
At use time, the model can replace its labeled reference patients with labeled patients from a new cohort. Its weights do not change. The local examples change instead.
That is the shift in the headline. Adapting to a new hospital would no longer necessarily mean building a new version of the model. It would mean assembling a local set of patients whose labels can act as the model’s evidence.
On the ADNI test set, the authors report a macro-AUC of 82.21% for classification. In external tests, they used the ADNI-trained model without updating its weights, then replaced its reference set with examples from OASIS-3 or SCAN. The reported macro-AUC was 0.742 on OASIS-3 and 0.808 on SCAN. The authors also report that mild cognitive impairment was the hardest class to distinguish in these external diagnosis tests.
These are retrospective public-cohort results. They do not establish diagnostic accuracy, safety, or benefit in routine care.
A clinic with a local memory of its patients
If the result holds in prospective clinical settings, a radiologist in a memory clinic could open a patient’s MRI and routine history beside a small, approved set of previously assessed local patients. The system could show which reference cases most influenced its estimate instead of sending the scan through a newly retrained model.
That could make a prediction easier to inspect: not only a fixed score, but a comparison with identifiable patients from the same clinical environment. It could also offer an alternative to repeated fine-tuning when a hospital’s population or records change.
The future depends on less glamorous work. A clinic would need representative local labels, compatible imaging and data processing, and validation for the intended use. Missing clinical variables would also have to be handled reliably. Until those conditions are met, the technology shifts the hard part of Alzheimer’s AI; it does not remove it.
