Donepezil for MCI: What a Three-Year Trial Actually Found, Buried Inside Its Own Subgroup
A patient with mild cognitive impairment, well-informed and proactive, asks to start donepezil preemptively. The largest trial ever built to answer exactly this question returned a negative headline result — with one real, specific exception hiding inside it.
Dr. Nadia S., 67, retired two years ago from a career as a hospital pathologist and has approached her own recent memory concerns with the same rigor she once brought to a difficult biopsy — she requested her own neuropsychological testing after noticing she'd started re-reading pages without registering them, tracked her symptoms in a spreadsheet for eight months before her appointment, and arrived today having already read the major trial literature on mild cognitive impairment. She is otherwise healthy, still swims most mornings, and has watched this particular disease more closely than most patients ever will: her own mother spent the last six years of her life with Alzheimer's disease, an experience Nadia has said directly she does not want to repeat unprepared. Testing confirms amnestic MCI: memory performance below age-expected norms, with other cognitive domains and daily functional independence both preserved.
She has also, on her own initiative, obtained APOE genotyping through a direct-to-consumer service: she carries one copy of the ε4 allele. She asks today, directly and specifically, whether she should start donepezil now, preemptively, rather than wait for progression to dementia. It is, in a specific sense, exactly the right question to bring evidence to bear on: the largest randomized trial ever built to test this precise scenario — donepezil in amnestic MCI, followed for three years, powered to detect delayed progression to Alzheimer's disease — already exists. Its headline result was negative. But its own genotype-stratified subgroup analysis, the same one Nadia's do-it-yourself testing puts her squarely inside, found something the topline result alone doesn't capture.
In clinic, discussing the workup results
I'd tell you plainly, Dr. S., that donepezil isn't an established therapy for MCI. The trial built specifically to test this — Petersen and colleagues, New England Journal of Medicine, 2005, 769 patients with amnestic MCI randomized to donepezil, vitamin E, or placebo, followed for three years — found donepezil reduced the risk of progressing to Alzheimer's disease at one year and two years, but that difference wasn't sustained at three years, the trial's actual primary endpoint.
That pattern is consistent with what donepezil's mechanism predicts: symptomatic cholinergic enhancement can mask underlying decline temporarily without truly slowing it, which is a different thing from disease modification, and would produce exactly this shape of early-then-fading effect.
I'd add the piece of that trial that doesn't show up in the topline result. Genotype was a pre-specified stratification variable in that trial, and in the hazard-ratio analysis, patients carrying at least one APOE4 allele — which you are — showed a benefit that held through all three years, hazard ratio 0.66, p equals 0.039 as originally reported.
The overall three-year result going negative doesn't erase that — it partly reflects the non-carriers in the trial, who didn't show a sustained effect, diluting the average. For a patient who actually has this genotype, that specific subgroup result is a real, if narrower, piece of evidence worth weighing directly.
I want to name two things about that finding, not to dismiss it, but to place it correctly. First, the paper itself reports that once the hazard-ratio analyses were corrected for multiple comparisons, the three-year APOE4 result no longer reached significance — p rose to 0.078. Those analyses were secondary, and the authors said plainly they hadn't adjusted the published p values. Second, and more generally: a subgroup result that wasn't the trial's primary comparison is hypothesis-generating, not something I'd act on by itself without replication — that's true regardless of how plausible the biology sounds, and it would be true even if the subgroup cut the other way.
Given that you already know your genotype, I'd rather use that information differently: a referral to discuss amyloid-PET and eligibility for the newer anti-amyloid antibody trials makes more direct use of your APOE4 status than restarting a decades-old symptomatic drug on a secondary finding its own trial didn't set out to prove.
Agreed: no donepezil started today. Nadia is referred for amyloid-PET and anti-amyloid antibody eligibility evaluation given her known APOE4 status, with structured cognitive re-testing at six months regardless of that referral's outcome.
Not agreed, and left explicitly open: whether the APOE4-subgroup finding from Petersen 2005 should be revisited as a live option if the anti-amyloid pathway turns out not to be available or appropriate for her. The psychiatrist views it as a real, if secondary, piece of evidence to keep in reserve; the pharmacologist maintains that an unreplicated subgroup result shouldn't become a fallback plan simply because a preferred option falls through.