Six stages, one number: a bid to make Alzheimer's prognosis legible

A Korean team merged blood tests, MRI, amyloid scans and cognitive status into a single six-stage prognostic ladder, then tested it in an American cohort. The early rungs held. The top ones had almost nobody standing on them.

The question that follows a diagnosis is almost never "what is it". It is "how fast". Families ask it in the corridor, after the appointment, when the word Alzheimer's has already been said and what they actually want is a timetable. Clinicians rarely have one to give, because the labels in widest use, cognitively unimpaired, mild cognitive impairment, dementia, describe where someone is standing, not how quickly the ground is moving. Two people with the same label can travel at very different speeds.

A study published in Nature Communications on February 2, 2026 tries to turn that speed into something you can name. Daeun Shin, Sang Won Seo, Kyunga Kim and colleagues used 1,263 participants from K-ROAD, the Korea-Registries to Overcome dementia and Accelerate Dementia research cohort, made up of 224 cognitively unimpaired people, 779 with mild cognitive impairment and 260 with dementia, 62.5 per cent of them women. Amyloid PET positivity climbed across those three groups, from 36.2 per cent to 59.8 per cent to 84.2 per cent. Rather than sorting people by symptoms alone, the team fed clinical risk factors, plasma markers and imaging measures into survival models and asked which combinations actually predicted worsening, defined by thresholds on the clinical dementia rating sum of boxes: 3 for very mild dementia, 4.5 for mild, 9.5 for moderate. Progression was tracked over a median follow-up of 2.1 years (95 per cent confidence interval 2.0 to 2.2), with a median of only two CDR assessments per person.

What the models picked is the interesting part, because it changes with where in the illness you look. Among cognitively unimpaired participants, the strongest prognostic contributor was plasma GFAP, a marker of astrocyte activation. In mild cognitive impairment it was hippocampal volume on MRI. In dementia it was age, and younger age at that, consistent with the long-observed pattern that earlier-onset Alzheimer's tends to move faster. Plasma phosphorylated tau-217, the blood marker that has dominated the field's attention, was not the single dominant predictor anywhere, but supplied consistent secondary information across the whole span. No one marker rules the disease; which instrument matters most depends on the stage you are already in.

Out of that came a unified ladder of six stages, labelled 0 through IVA and IVB. In K-ROAD, every adjacent pair of stages differed significantly in progression risk, and higher stages tracked with rising CDR-SB and falling MMSE scores. The authors also report inflection points, places where incidence of decline jumps rather than drifts, at the transitions from 0 to I, I to II, II to III, and III to IVA. Lower stages were separated mainly by the very mild and mild dementia outcomes; the top two stages were driven by progression to moderate dementia.

Then came the harder test. The team applied the same staging to the Alzheimer's Disease Neuroimaging Initiative in the United States: 290 participants, 160 cognitively unimpaired, 118 with mild cognitive impairment, 12 with dementia, followed for a median 5.5 years. Separation held for very mild dementia (log-rank P = 2.03 x 10 to the minus 6) and mild dementia (P = 3.52 x 10 to the minus 5), and was weaker for moderate dementia (P = 0.0419). But stages III, IVA and IVB had 10 or fewer participants each and were dropped from that analysis altogether. Only 290 of 378 eligible ADNI participants had the full set of measures required, which left a biomarker-enriched sample thin on advanced dementia. The authors say plainly that broader, population-based validation is needed, and that the framework is not built for treatment decisions: eligibility for anti-amyloid antibodies turns on confirmed amyloid positivity and response biomarkers this system does not assess.

For a family, the honest reading is that this is scaffolding, not a verdict. If it survives testing in more varied populations, it could make the sentence "your risk of decline over the next few years is in this band" mean the same thing in Seoul as in Sacramento, and it could let trials enrol people who are actually moving at comparable speeds. That is worth having. What it cannot yet do is tell any individual person how much time they have.

The paper's own framing is the tell: an exploratory framework, offered to complement biological staging rather than replace it. A ladder with well-measured lower rungs and almost no one tested on the top three is still, for now, a ladder to be checked before it is climbed.