Exercise Beats Donepezil’s Modest Benchmark in Alzheimer’s

Exercise has a better cognitive signal than donepezil in a major indirect comparison of randomized trials. That finding says as much about the low bar set by older Alzheimer's drugs as it does about exercise. Donepezil produced a standardized MMSE effect of Hedges' g=0.23 versus placebo in Alzheimer's disease; exercise produced g=0.46 in Alzheimer's disease and g=0.63 in mild cognitive impairment (MCI). Those are useful comparisons, but they are not head-to-head trials, and they do not turn a gym membership into Alzheimer's treatment.

EPA-dominant omega-3 supplements make the picture stranger. A 2023 network meta-analysis ranked high-dose EPA first, with an SMD of 3.00, yet that estimate came from only six omega-3 trials and conflicts with broader reviews of supplementation trials. Exercise has the cleaner case. EPA has a conspicuous number that needs replication before anyone treats it as a cognitive prescription.

Donepezil's benchmark is modest

Donepezil remains a common symptomatic treatment for Alzheimer's disease, but its average cognitive effect is small. In Pisani et al.'s 2021 meta-analysis, 18 donepezil randomized controlled trials totaling 2,984 participants with Alzheimer's disease produced an MMSE effect of g=0.23 and an ADAS-Cog effect of g=−0.17 versus placebo. On the 30-point MMSE, the authors' methodological conversion corresponds roughly to 1.1 to 1.2 points in a typical Alzheimer's trial population.

A statistically detectable difference from placebo does not automatically deliver a large day-to-day change for a patient or caregiver. The available evidence also found no detectable cognitive benefit from donepezil in MCI. That distinction matters: MCI is not Alzheimer's dementia, and evidence from one group cannot be casually carried into the other.

Exercise compares well, with an important asterisk

The same meta-analysis pooled 19 exercise trials: 524 participants with Alzheimer's disease and 1,269 with MCI. Exercise improved MMSE scores relative to control conditions by g=0.46 in Alzheimer's disease and g=0.63 in MCI. On that one cognitive scale, the Alzheimer's estimate was about twice donepezil's, while the MCI estimate was larger still.

The comparison is indirect. The exercise and donepezil trials differed in participants, control conditions, blinding, placebo quality, and co-interventions. Drug trials can use placebos; exercise trials cannot fully blind participants to whether they are exercising. The exercise evidence in Alzheimer's disease also came from a much smaller sample than the donepezil evidence: 524 versus 2,984 participants. Those differences prevent a claim that exercise has definitively beaten donepezil.

Still, the result fits a broader literature. A 2024 Bayesian network meta-analysis by Yuan Yuan et al. found aerobic exercise and combined aerobic-plus-resistance exercise ranked highest for cognition in Alzheimer's disease and dementia populations. It proposed reduced beta-amyloid deposition and better cerebrovascular function as candidate mechanisms, but those biological findings do not prove that either pathway caused the cognitive effects.

Other reviews point in the same direction without settling the best program. A 2024 systematic review and meta-analysis linked long-term exercise and multidomain interventions with better cognitive outcomes and lower risk of MCI and dementia in older adults. A 2025 network meta-analysis likewise examined which exercise types, frequencies, and durations best support cognition in older adults. These results support activity as a serious brain-health intervention, but they do not establish one universal dose for people with MCI or Alzheimer's disease.

The 2025 EXERT study is especially relevant because it was a multicenter phase 3 randomized trial of supervised exercise in adults with MCI, measuring cognition and Alzheimer's biomarkers. The EXERT trial strengthens the case for studying structured exercise in this population, but its available summary does not provide enough outcome detail to quantify the cognitive or biomarker results here.

EPA's enormous estimate has a small foundation

Tseng et al.'s 2023 network meta-analysis compared 52 randomized trials, totaling 21,111 participants, of FDA-approved therapies and nutritional interventions in Alzheimer's dementia. It ranked EPA-dominant omega-3 supplementation—about 1,500 to 2,000 mg daily combined with antioxidants—first for cognitive function, with SMD=3.00 (95% CI, 1.84 to 4.16).

That is a striking result. It is also a reason to slow down. Only six omega-3 trials fed the EPA node. The confidence interval spans 2.32 standardized mean-difference units, which signals substantial imprecision even though it excludes zero. The published abstract does not report heterogeneity for those six trials, so readers cannot judge how consistent their results were. The acceptability estimate was similarly unstable: discontinuation odds ratio 0.46, with a 95% CI of 0.04 to 5.87. That interval is compatible with far fewer or far more discontinuations and cannot guide tolerability decisions.

The larger problem is disagreement with the rest of the omega-3 literature. Kosti et al.'s 2022 systematic review and dose-response meta-analysis found that EPA+DHA supplementation trials had failed to provide evidence of cognitive protection. These studies used mixed EPA/DHA formulations rather than the EPA-dominant regimen in Tseng et al., so they do not directly refute that exact intervention. But they make a huge EPA estimate less believable until dedicated, larger trials reproduce it.

Observational research offers no clean rescue. Wei et al.'s 2023 review of prospective cohorts found inconsistent associations among omega-3 supplements, dietary intake, blood markers, dementia, and cognitive decline. Diet quality, exercise, income, education, and healthcare access can all travel with fish intake. Cohort associations can generate hypotheses; randomized supplementation trials must carry the causal claim.

Newer drugs could change the comparison

Donepezil is not the entire drug story. Anti-amyloid antibodies such as lecanemab and donanemab target disease pathology rather than acetylcholine signaling. A 2025 network meta-analysis of updated Alzheimer's drugs described encouraging cognitive and clinical results, while noting continued controversy. Its available abstract does not provide MMSE or ADAS-Cog standardized effects that can be compared fairly with the g=0.23 donepezil estimate or the exercise estimates.

That missing comparison matters. It would be wrong to use a 1990s-era symptomatic drug as the sole benchmark for all modern pharmacotherapy. It would be equally wrong to assume newer drugs have already established a cognitive advantage over exercise without directly comparable data.

What to do with this evidence

Exercise deserves action now because several randomized-trial syntheses support cognitive benefit, its broader physical benefits are well established, and its risks are manageable when the program fits the person's health and mobility. Aerobic and combined aerobic-plus-resistance programs have the best mode-specific support in Alzheimer's disease and dementia evidence, but clinicians or qualified exercise professionals should adapt activity for falls risk, cardiovascular disease, frailty, pain, and disability.

EPA does not deserve the same confidence. The SMD=3.00 result is interesting, not practice-changing: six trials, no reported heterogeneity in the available abstract, an uninformative discontinuation estimate, and conflict with broader supplementation evidence. No human randomized trial in this evidence base has tested exercise plus EPA against either intervention alone or against Alzheimer's pharmacotherapy. Claims that the combination works through complementary mechanisms remain theory.

The practical conclusion is less glamorous than "lifestyle beats drugs." Exercise has a credible cognitive signal and many independent reasons to recommend it. Donepezil's average effect is modest. EPA remains a replication problem wearing a very large effect size.