One nun lived to 101, sharp and articulate to the end. At autopsy her brain was packed with Alzheimer's pathology.
Most people carry exactly two thoughts about this disease: "I'll deal with it later" and "it's luck of the draw." The best evidence of the last decade unsettles both — how much pathology piles up in your brain and how severe your symptoms are can come apart, and part of the buffer in between is in your hands. This one is mechanism only: when the damage starts accumulating in silence, why accumulation doesn't guarantee symptoms, how many levers there genuinely are, and which of the tests promising to "catch it twenty years early" are real progress.
In 1986 an epidemiologist recruited 678 Catholic nuns and followed them for the rest of their lives: annual cognitive tests, and their brains donated after death. It's the sample researchers dream of — same schedule, same food, same community, which flattens a mountain of confounders for free.
One of them lived to 101 and was still scoring beautifully on cognitive tests shortly before she died. But open her brain and it was full of plaques and tangles — the two signature lesions of Alzheimer's: protein junk clumping outside the cells, and tangled filaments knotting up inside them. By pathology she was a severe case. By her own behaviour she wasn't.
She isn't a fluke. Plenty of older people turn out at autopsy to have enough pathology to meet diagnostic criteria while never having been demented in life. Hence a dedicated concept: cognitive reserve.
Two things need separating. Brain reserve is hardware quantity — how big the brain is, how many neurons and synapses there are. Cognitive reserve is software flexibility — whether the network can still get the same job done by another route. An analogy: it's not that the main road is wider, it's that the city has more side streets and alleys, so when a main artery goes down traffic reroutes and from the outside everything still flows.
More years of education, complex work, bilingualism, an active social life, still learning things in old age — all correlate with higher reserve. To be clear about the limits: this is mostly observed correlation, not causation from randomised groups.
There's a counterintuitive flip side too: once high-reserve people do become symptomatic, they tend to decline faster. They were coping by rerouting, and held out until the damage was higher before being overwhelmed — by which point there are few routes left, so the collapse is a cliff. Reserve buys time, not immunity. (The disease hits the entorhinal cortex and hippocampus first — exactly the tissue that writes new experience into memory: hippocampus & entorhinal cortex)
Artificial networks give you the isomorphism for free: pruning tolerance. Randomly delete a fraction of a trained network's connections and performance often barely moves — because it uses distributed representations: a concept doesn't live in one unit, it's smeared across the joint activity of thousands, so knock some out and the rest still reconstruct it. Past some fraction, though, performance falls off a cliff. That curve's shape — nothing noticeable for a long while, then sudden collapse — is exactly what the clinic sees with cognitive reserve.
Meet the two kinds of junk. Aβ (beta-amyloid) is a fragment enzymes snip out of a normal membrane protein; snipped the wrong way it's unusually sticky, and enough of it clumps into plaques outside the cell. Tau is normally the sleeper on the neuron's internal "rails", keeping the transport tracks stable; over-modified, it falls off and knots into tangles, and the tracks go with it.
The amyloid cascade hypothesis, proposed in 1992, says Aβ is the trigger at the very top: pull it, and tau, synaptic death and finally symptoms follow in order. That story ruled for thirty years.
Then it started leaking. Leak one: plaque load tracks symptom severity poorly — see the nun above; tau tangle distribution, on the other hand, tracks it rather well. Leak two was worse: dozens of Aβ-clearing drugs failed, one after another, over two decades. Around 2020, "the amyloid hypothesis is dead" was close to a catchphrase.
The turn came in the last three years. Lecanemab (full FDA approval in 2023) slowed decline on a cognitive-functional scale by roughly 27% over an 18-month phase 3 trial; donanemab was approved in July 2024, pointing the same way. For the first time in thirty years, clearing Aβ actually changed the course of the disease.
Don't take either extreme on this result. Looking back, it proves Aβ really is on the causal chain — an innocent bystander wouldn't produce any effect when removed. Looking forward, the effect size is sobering: that 27% is an absolute difference of 0.45 points on an 18-point scale, which many clinicians doubt patients and families can feel. The cost isn't trivial either: these antibodies cause ARIA (brain swelling and microbleeds) — roughly 12.6% and 17.3% in the lecanemab arm, higher for donanemab (about 24% and 31.4%) — and carriers of one particular genotype are at greater risk.
The real lesson is on the timeline: Aβ starts accumulating ten to twenty years before symptoms. By the time you notice your memory slipping, the trigger has long been pulled and tau and synapse loss are running under their own steam. The drug hits the top of the cascade at the very bottom of the timeline — a small effect is structural. Which is why the whole field is now pushing earlier.
Something else textbooks tidy up too much: pure Alzheimer's pathology is the minority in older brains. What autopsies keep finding is mixtures — plaques and tangles layered on top of small-vessel disease and other protein deposits. So "which disease do I have" is already the wrong question.
The vascular line is the most actionable. The brain is 2% of body weight and burns a fifth of the body's energy, supplied by an extraordinarily dense capillary net. Endothelium, pericytes and the astrocytic "endfeet" wrapped around them form the neurovascular unit: it delivers glucose and oxygen and doubles as the barrier keeping junk in the blood out. Midlife hypertension, high LDL and diabetes grind down exactly that plumbing — the long-range wiring in white matter degrades under chronic under-perfusion. "What's good for the heart is good for the brain" isn't a slogan; it's the same pipes.
Clearance is the other line. Aβ can pile up because production is high or because removal is slow. Removal runs through several routes: enzymes degrading it, fluid flow along the outside of blood vessels washing it out, and microglia eating it. One unsettled controversy has to be flagged here: work since 2013 proposed that this fluid clearance is enhanced during sleep, and human studies found Aβ rises after a single night of deprivation — but a 2024 study using a new method reported the opposite (clearance falling during sleep), and was then challenged in the same journal on the grounds that its data don't support that conclusion. This is still live; don't treat it as settled (Topic 32 takes it apart).
Finally, those janitors. Microglia are the brain's own immune cells, patrolling the tissue and eating debris. But they have a second face: during development they use a "complement" tagging system to prune surplus synapses — tagged connections get eaten, and this is a necessary step in shaping the brain. The trouble is that the program can be wrongly restarted in a degenerating brain, taking connections that should have stayed; and one known Alzheimer's risk gene happens to encode a receptor on the microglial surface. The janitors and the arsonists are the same cells. (What each glial type actually does: glial cells)
What microglia do has a name in AI: pruning — deleting the "unimportant" connections in a trained network to save compute at almost no cost in performance. Everything hinges on the importance score: the algorithm uses some metric to decide what's deletable, and if the metric is wrong, what gets cut is a critical pathway. Microglia score with complement tags — "this synapse is quiet, unused: tag it, clear it." In development that's a bargain; in old age, wrongly triggered by pathology, the same algorithm prunes connections still in use. The problem is never pruning itself, it's whether the score is right.
In July 2024 the Lancet Commission on dementia published the most authoritative list currently available: 14 modifiable risk factors, linked in theory to around 45% of dementia cases. Two are new relative to the 2020 edition — midlife high LDL cholesterol, and uncorrected vision loss.
Read the 45% correctly first. It's a population attributable fraction: the share of cases that would theoretically disappear if all 14 factors were eliminated across the whole population and every association were causal. It is not "do these things and cut your own odds by 45%."
And "every association is causal" is exactly the brittle part. Depression, staying in, moving less — these may well be prodromal signs of dementia rather than its causes; the pathology was already moving, and the person merely lost the appetite for company first. Observational research struggles to scrub out that reverse causation.
So look at randomised trials. Two examples: hearing — a large 2023 trial that fitted older adults with hearing aids found no difference in the population overall after three years; but in a prespecified high-risk subgroup, cognitive decline was 48% slower. Discount subgroup results, yes — but this one was declared in advance, and it's worth treating as a lead. Lifestyle — in a 2025 US trial of over two thousand people, a structured multidomain intervention (regular exercise + a specific diet pattern + cognitive training + social engagement + blood-pressure monitoring) produced better global cognition than a self-guided control, replicating the direction of the earlier Finnish trial in a more diverse population. The effect is modest, but it came from randomisation, which makes it worth more than a pile of observational studies.
Now early detection. In May 2025 the FDA cleared the first Alzheimer's blood test — a plasma ratio of one specific form of tau to Aβ, with high agreement against the gold-standard brain scan (about 91.7% positive, 97.3% negative). That's genuine progress: the alternatives were a lumbar puncture or an expensive PET scan.
But the indication is written with restraint: it's for differential diagnosis in people who already have cognitive symptoms, not a fortune-teller for the well. Testing asymptomatic people — blood test or risk gene — isn't currently recommended, for practical reasons: a positive result carries no definite action (the things worth doing are worth doing either way); a false positive costs years of real anxiety; and the insurance and career consequences are irreversible. If you do have persistent cognitive problems affecting daily life, it's the opposite: go to a memory clinic, where the blood test is valuable — it helps separate Alzheimer's from depression, thyroid problems and drug side effects, all reversible causes, and that last category is bigger than most people assume. Meanwhile the "brain health panels", coconut oil and memory supplements on the market are a different story — unsupported.
Honestly: the things with the hardest evidence behind them are all boring — control blood pressure, correct hearing and vision, move, don't smoke, don't shut yourself in, wear a helmet. Not sexy. But the only ones with randomised trials behind them.
This disease turns a philosophical puzzle into a daily problem on the ward: when memory is peeled away layer by layer, is what remains the same person?