• Dates: 2026-09-19–present
  • Status: unfinished
  • Importance: 5

My friend Doug went from comp bio PhD student to developing drugs for Alzheimer’s and schizophrenia. I was inspired to ask: would it make sense for me to jump into dementia research at some point? I talked with GPT, and here’s my simplified take.

Approaches to dementia

Dementia affects tens of millions of people worldwide: the WHO estimates that 57 million people were living with it in 2021. My grandma suffered from dementia in her final years. My working conception of Alzheimer’s, the most common type of dementia, is a disease where plaques spread through the brain, neurons die, and neuronal connections are disrupted.

There seem to be a few broad approaches to reducing this burden.

Prevention

The Lancet 2024 meta-analysis seems to be the latest canonical work on this. It names and ranks 14 modifiable factors that could be related to dementia. I’ve grouped them into four categories:

  • Sensory loss/injury: hearing loss, untreated vision loss, traumatic brain injury.
  • Cardiovascular and metabolic issues: high LDL cholesterol, physical inactivity, diabetes, hypertension, obesity.
  • Cognitive/psychological/social factors: social isolation, less education, depression.
  • “Pollutants”: air pollution, smoking, excessive alcohol consumption.

How many cases of dementia could be prevented by modifying these factors? The report estimates about 45% could theoretically be prevented or delayed, assuming the associations are causal and the risk factors could be eliminated.1

I take this number with fistfuls of salt. The 45% came from a methodology I didn’t understand, plus we don’t know how causal these factors are and whether the field has looked at the right factors. But the upshot for me is that plausibly a decent fraction of cases are preventable, by doing things you would do anyway for your overall health:2

  • Sensory loss/injury: Get appropriate hearing and vision care, and take precautions against head injuries when driving, biking, or playing sports. (However, the evidence for hearing aids for dementia is weak right now.3)
  • Cardiovascular and metabolic health: Exercise, eat well, and manage blood pressure, cholesterol, and diabetes.
  • Cognitive/psychological/social health: Spend time with family and friends, stay mentally active, and seek care for depression.
  • “Pollutants”: Avoid smoking, limit alcohol, and reduce exposure to air pollution where possible.

Thus, existing efforts on these fronts may also help reduce dementia down the road.

Medicine

Alzheimer’s currently has no cure; the best results are a 27% slowing of cognitive decline after 18 months:4

Treatment Trial and duration Slowing of decline versus placebo
Lecanemab CLARITY AD; 18 months 27% on the Clinical Dementia Rating–Sum of Boxes (CDR-SB), a measure of cognition and daily function.5
Donanemab TRAILBLAZER-ALZ 2; 76 weeks 22% on the integrated Alzheimer Disease Rating Scale (iADRS) in the combined population; 35% in the prespecified group with low or medium tau levels.6

A big bottleneck on progress is the slowness of clinical trials — they used to last 1-2 years (it takes a while for people with mild cognitive impairment to progress to fuller dementia) and the latest ones are 4-6 years as the field moves to treat people before they even show symptoms. The main thrust of Alzheimer’s treatment for the past three decades has been to neutralize and remove amyloid beta, a protein subunit of the aforementioned amyloid plaques which have been observed at higher levels in people with dementia:

NIH illustration showing beta-amyloid accumulating into plaques between neurons.

Illustration of beta-amyloid plaques, not a microscope image. Credit: National Institute on Aging, National Institutes of Health. Source and public-domain reuse information.

Amyloid-targeting treatments

One major thrust of Alzheimer’s drug development has been to reduce amyloid-beta, the peptide that aggregates into plaques. Researchers have tried vaccines, manufactured antibodies, and drugs that inhibit amyloid production. There’s a short GPT-written history supplement below.

The striking part for me is that lecanemab and donanemab substantially reduced amyloid, yet still provided, at best, only the 27% “slowing” clinical benefit shown above. They also increased the risk of amyloid-related imaging abnormalities (ARIA): swelling or fluid accumulation (ARIA-E), and small hemorrhages or deposits left by bleeding (ARIA-H).56

In the pivotal trials, ARIA-E occurred in about 13% of lecanemab recipients and 24% of donanemab recipients; symptomatic ARIA-E occurred in about 3% and 6%, respectively. Most imaging abnormalities caused no symptoms, but serious and sometimes fatal events can occur.7

One possible explanation for the modest benefit is timing. Alzheimer’s pathology can begin 10–20 years before symptoms, so treating someone after cognitive problems appear may leave substantial damage already in place. The AHEAD investigators explicitly give this as a reason to test earlier intervention. It’s a plausible hypothesis but remains to be shown in the main ongoing clinical trials:8

Study Treatment Estimated primary completion Estimated overall completion
AHEAD 3-45 Lecanemab December 21, 2028 January 16, 2031, including the extension
TRAILBLAZER-ALZ 3 Donanemab November 2027 November 2027

Other medicines

Doug tells me that his mom mainly sees Alzheimer’s patients who still take donepezil rather than these newer antibodies. Donepezil was first approved in the US in 1996. It inhibits acetylcholinesterase, increasing the availability of the neurotransmitter acetylcholine; it can help symptoms, but hasn’t been shown to alter the underlying disease process.9

I can see why a familiar oral medication might remain common when the newer treatments involve more monitoring and potentially serious adverse effects. For me, this is a reminder to keep an open mind about where a much better treatment might come from. Knowing a drug’s molecular target doesn’t mean we fully understand its effects on a complicated disease. Even the seemingly simple idea of “remove amyloid” has taken decades to turn into modest clinical benefit. Doug’s bet is on making donepezil analogs.

Other approaches

What about brain training? There are intriguing results. In a ten-year analysis of the randomized ACTIVE trial, a specific speed-of-processing training program was associated with a 29% lower hazard of dementia; memory and reasoning training did not show statistically significant reductions.10

I’d want to distinguish getting better at a trained task, improving everyday functioning, and actually preventing dementia. Those are different outcomes. The National Institute on Aging cautions that the ACTIVE findings don’t establish equivalent benefits for commercially available brain-training apps. This is an area I’d like to understand better.

Does it make sense for me to work on this?

My impression is that Alzheimer’s is already a crowded research field. At the same time, going after it still feels very noble (and anyone who does it is worth celebrating).

I’m hopeful about the ongoing trials. But the importance of the problem doesn’t by itself tell me that switching into Alzheimer’s drug development would be my best contribution. I’d want a specific opportunity where my skills could add something useful.

Never say never. An opportunity to apply AI might be a good fit, or perhaps there’s useful work in helping people adopt interventions such as hearing care. For now, though, while I may keep learning and writing about dementia, it probably makes more sense for me to spend my free time on health AI.

Supplement: A short history of amyloid-targeting treatment

First, a striking result in mice. In 1999, Dale Schenk and colleagues reported that immunizing genetically engineered mice with amyloid-beta could prevent or reduce Alzheimer’s-like plaque pathology. It was an exciting demonstration that the immune system could act on the deposits. It wasn’t yet evidence that the same approach would preserve human cognition.11

Then, trouble in humans. The AN1792 vaccine program tested that idea in people with mild-to-moderate Alzheimer’s. Dosing in its phase IIa trial was stopped after about 6% of immunized participants developed meningoencephalitis, inflammation of the brain and its surrounding membranes. A promising target had collided with the difficulty of controlling an immune response.12

There was more than one way to pursue amyloid. Researchers also tried reducing its production. Verubecestat, for example, inhibited the enzyme BACE1. But a trial in people with prodromal Alzheimer’s found no clinical benefit, and some measures favored placebo. Changing amyloid biology was not automatically beneficial.13

Eventually, antibodies showed clinical benefit. Lecanemab and donanemab supplied the positive results described above, alongside substantial plaque reduction and the risk of ARIA. That is meaningful progress, but it leaves a large gap between removing a pathological feature and stopping the disease.

Now, the question is partly when—and how—to intervene. Earlier treatment has already had one major disappointment: the A4 trial found that solanezumab did not slow cognitive decline over 240 weeks in cognitively unimpaired people with elevated amyloid. That antibody targeted soluble monomeric amyloid and did not reduce plaque burden below baseline. AHEAD and TRAILBLAZER-ALZ 3 test different antibodies, so A4 doesn’t settle their prospects. But it does mean that “treat earlier” remains a hypothesis to test, not an answer we already have.14

  1. Livingston et al., Dementia prevention, intervention, and care: 2024 report of the Lancet standing Commission, especially Table 1 and the “New PAF calculation” section. The estimate is 45.3%, rounded to 45%. The calculation combines risk-factor prevalence and relative risk, then adjusts for overlap among factors using “communality” estimates from the Norwegian HUNT cohort. I understand the motivation for avoiding double-counting, but haven’t independently validated this adjustment. Open-access manuscript and appendix. ↩

  2. “Modifiable” doesn’t mean entirely under an individual’s control. Genetics, education, income, access to care, and the surrounding environment all constrain what changes are feasible. This isn’t a claim that people who develop dementia could simply have chosen otherwise. ↩

  3. Lin et al., ACHIEVE trial, The Lancet (2023). The trial randomized 977 adults aged 70–84 with untreated hearing loss to hearing aids plus audiological counseling or a health-education control. After three years, there was no significant difference in cognitive decline in the full sample. In a prespecified analysis of the 238 participants recruited from the ARIC cardiovascular cohort, who were at higher risk of cognitive decline, hearing intervention slowed decline by about 48%. The other 739 participants showed no significant benefit. That 48% refers to cognitive-score decline in a subgroup, not a 48% reduction in dementia diagnoses. ↩

  4. Clinical studies measure this as follows: the control and treatment groups are given cognitive tests at baseline and after say 18 months; the decline in scores in the treatment group is x% less than in the control group. ↩

  5. Van Dyck et al., Lecanemab in Early Alzheimer’s Disease, NEJM (2023). CLARITY AD enrolled people with amyloid-confirmed mild cognitive impairment or mild dementia due to Alzheimer’s. ↩ ↩2

  6. Sims et al., Donanemab in Early Symptomatic Alzheimer Disease: The TRAILBLAZER-ALZ 2 Randomized Clinical Trial, JAMA (2023). The exact iADRS estimates were 22.3% in the combined population and 35.1% in the low/medium-tau population. On CDR-SB, the corresponding estimates were 28.9% and 36.0%; the percentage depends on the population and outcome measure. ↩ ↩2

  7. In CLARITY AD, ARIA-E occurred in 12.6% with lecanemab versus 1.7% with placebo; ARIA-H occurred in 17.3% versus 9.0%. In TRAILBLAZER-ALZ 2, ARIA-E occurred in 24.0% with donanemab versus 2.1% with placebo; ARIA-H occurred in 31.4% versus 13.6%. These categories overlap and should not be added. See the trial reports above. Donanemab’s recommended titration was subsequently changed: its prescribing information reports lower ARIA-E with the newer schedule and warns that serious, life-threatening, and fatal ARIA can occur. ↩

  8. Rafii et al., The AHEAD 3–45 Study: Design of a prevention trial for Alzheimer’s disease, Alzheimer’s & Dementia (2023; published online in 2022). The investigators describe pathology beginning 10–20 years before symptoms and the rationale for intervening before substantial downstream damage. ↩

  9. Aricept prescribing information, particularly “Initial U.S. Approval” and section 12.1, “Mechanism of Action.” ↩

  10. Edwards et al., Speed of processing training results in lower risk of dementia, Alzheimer’s & Dementia: Translational Research & Clinical Interventions (2017). ACTIVE originally randomized 2,802 older adults. This analysis reported a dementia hazard ratio of 0.71 (95% CI 0.50–0.998) for speed training versus control. Dementia was identified through a combination of cognitive/functional measures and reported diagnoses, rather than a uniform clinical diagnostic examination of everyone. A 2026 follow-up using Medicare claims also reported a signal among speed-training participants who completed booster sessions (HR 0.75, 95% CI 0.59–0.95), but not those without boosters. That result should not be generalized to everyone assigned to speed training. ↩

  11. Schenk et al., Immunization with amyloid-beta attenuates Alzheimer-disease-like pathology in the PDAPP mouse, Nature (1999). ↩

  12. Gilman et al., Clinical effects of A-beta immunization (AN1792) in patients with AD in an interrupted trial, Neurology (2005). ↩

  13. Egan et al., Randomized Trial of Verubecestat for Prodromal Alzheimer’s Disease, NEJM (2019). ↩

  14. Sperling et al., Trial of Solanezumab in Preclinical Alzheimer’s Disease, NEJM (2023). ↩