Field notes on the model
About AURORA
AURORA — Alzheimer's Unified Representation for Omni-modal Risk Analysis — is a multimodal research model for Alzheimer's disease analysis. It combines clinical information with MRI-based information to identify patterns associated with Alzheimer's disease. This page walks through what it is, how it was built, and — just as importantly — what it is not.
- Two models: XGBoost (a pattern-matching model for numeric test scores) on clinical scores, EfficientNetB1 (a small image-reading model) on MRI
- Trained on the ADNI research dataset
- Its limitations are published below, not left out
Jump to a section
In short: This is not a medical diagnosis. It cannot tell you whether you have Alzheimer's disease or any other condition.
AURORA is a research project, not a medical device, and nothing on this page changes that.
If a link brought you here before you'd picked a starting point, here are both:
Honest limitations
- AURORA has not been clinically validated, peer reviewed, or approved by any medical regulator. It was built as a science fair project.
- Performance varies by class and by data split. The imaging pathway is weak on its own and contributes mainly through fusion.
- The paired imaging-plus-clinical cohort is small, so the fusion weighting is tuned on limited data.
- Reported contributing factors are approximations of global feature importance, not per-patient attribution.
- ADNI participants are not representative of the general population, so results may not transfer evenly across groups.
AURORA and the memory check are different things
The at-home memory check does not run through AURORA. The check uses simple, transparent scoring of the questions and tasks you complete — because AURORA expects standardised scores administered by a clinician (MMSE, MoCA, Trail Making, and similar), and feeding it self-administered answers would produce a number nobody has validated.
The AURORA analysis is available separately, for people who already have clinical test results from a doctor and want to explore what a multimodal model makes of them.
How it's built
The technical detail below is for readers who want it — researchers, judges, anyone curious about the mechanism. Nothing past here changes what to actually do, which is the two links higher up this page.
What goes in
Clinical & neuropsych test scores
Standardised scores a clinician administers — things like MMSE, MoCA, and Trail Making — entered directly, not estimated from self-report.
MRI brain scan (optional)
Structural MRI slices. This input is optional — the tabular pathway alone can still produce a result without a scan.
Two separate pathways
Tabular pathwaygradient-boosted decision trees
A tree-based model (XGBoost) looks for patterns across the clinical and neuropsych scores — a standard, well-understood approach for this kind of structured data, and the stronger of the two pathways on its own.
Imaging pathwayEfficientNetB1
A convolutional neural network scans the MRI slices for patterns. It's deliberately a small, lightweight model so it can run on ordinary CPU hardware rather than needing a dedicated GPU — but it's also the weaker pathway on its own (see "Honest limitations" above), and contributes mainly in combination with the tabular pathway, not alone.
Combined
Weighted fusionweighted late fusion
The two pathways' outputs are combined using fixed weights, rather than being trained jointly end-to-end — a simpler, more inspectable way to combine them given how little paired imaging-plus-clinical data there is to tune a more complex combination on.
What comes out
CN / MCI / AD, with probabilities & key drivers
The final result is a classification across three categories — cognitively normal (CN), mild cognitive impairment (MCI), or Alzheimer's disease (AD) — a probability for each, and the features that most influenced that particular result.
Every stage above was trained on ADNI — the Alzheimer's Disease Neuroimaging Initiative, real patient MRIs and clinical records, the dataset widely used in Alzheimer's research. Tap any box above for the technical term and a plain-language explanation of it; nothing here changes what AURORA is for or what to actually do next, which is still the two links higher up this page.
What AURORA actually sees
The diagram above names the stages. This walks one fictional case through them, step by step, with the same layout a real analysis uses on /check or /tool — so you can see what actually moves between stages, not just that it does.
Step 1 of 5 — What goes in
A fictional case's clinical scores (synthetic)
None of these numbers belong to a real person. They're invented so this walkthrough has something concrete to carry through the pipeline below — standardised test scores, same as a clinician would enter.
Optional MRI: hippocampal volume rated "slightly reduced" by the fictional radiology note used for this example.