Cognitive health and dementia risk, what the modifiable-factor evidence shows

Dementia risk is often presented to the public in one of two incompatible ways: as something almost entirely determined by genetics and age, or as something a person can train away with the right app or puzzle habit. The actual evidence base sits between those two claims and is more interesting than either. This page separates the population-level modifiable-risk-factor estimate from the much narrower randomized evidence behind specific interventions, cognitive training programs chief among them.

The Lancet Commission’s modifiable-factor estimate

The most current, comprehensive synthesis of this evidence is the 2024 report of the Lancet standing Commission on dementia prevention, intervention, and care, led by Gill Livingston and an international panel of dementia researchers. The Commission identifies 14 modifiable risk factors across the life course, from early life through late life: less education, hypertension, hearing loss, high LDL cholesterol, depression, traumatic brain injury, physical inactivity, diabetes, smoking, excessive alcohol consumption, obesity, air pollution, social isolation, and untreated vision loss, with high LDL cholesterol and vision loss added as new factors in the 2024 update. Collectively, the Commission estimates these 14 factors account for approximately 45% of dementia cases worldwide, up from a 40% estimate across 12 factors in the Commission’s 2020 report.

That 45% figure is a population attributable fraction, a specific statistical calculation that combines how common each risk factor is in the population with how strongly each factor is associated with dementia risk in observational studies, then estimates what fraction of cases might not occur if the factor were eliminated entirely across the population. It is a rigorous and widely used method in public health epidemiology. It is not the same thing as a randomized trial demonstrating that removing all 14 factors from a given person’s life would cut that person’s own dementia risk by 45%; the underlying relative risks that feed into the calculation come predominantly from observational cohort studies, and the calculation itself assumes each factor’s association is causal and that the factors act independently of each other, assumptions the observational literature alone cannot fully confirm.

What a randomized multidomain trial actually found

The clearest randomized test of a multidomain risk-factor intervention is the FINGER trial (Ngandu and colleagues, Lancet, 2015), which randomized 1,260 at-risk older adults in Finland, aged 60 to 77 with an elevated dementia risk score, to a 2-year intervention combining diet, exercise, cognitive training, and vascular risk monitoring, or to a general-health-advice control group. The trial’s primary outcome was change in a composite neuropsychological test battery score, not a dementia diagnosis. The intervention group’s score improved by 0.022 points per year more than the control group’s (95% CI 0.002 to 0.042, p=0.03), a statistically significant but numerically modest between-group difference on a composite test score, over a study that ran 2 years, not the decades over which dementia typically develops. This is real randomized evidence that a multidomain intervention can modestly improve or maintain measured cognitive performance in an at-risk population; it is evidence about a test-battery score change over 2 years, not proof that the intervention prevented a single case of diagnosed dementia within the trial.

Brain training specifically, and why the transfer evidence is weak

Commercial cognitive-training products are frequently marketed with the implication that practicing on their exercises protects against age-related decline generally. The relevant randomized evidence does not support that broad a claim.

A Cochrane systematic review (Gates and colleagues, 2020), covering 8 randomized trials and 1,183 cognitively healthy older participants, examined computerized cognitive training lasting 12 weeks or more. Compared with active control conditions, the review found a slight improvement in global cognitive function immediately after training (standardized mean difference -0.31), but rated the certainty of this evidence low, and found no clear evidence the effect persisted 12 months later in the one trial that measured it. The review found little or no effect on episodic memory or working memory specifically, rated the overall quality of evidence for every outcome as low or very low, and concluded that “these benefits are of uncertain clinical importance.”

A longer-running trial adds a more specific pattern to this picture. The ACTIVE trial’s 10-year follow-up (Rebok and colleagues, 2014), following 2,832 older adults randomized to memory training, reasoning training, speed-of-processing training, or no-contact control, found that reasoning and speed-of-processing training maintained measurable effects on their own specific trained ability a full decade later (reasoning effect size 0.23; speed of processing effect size 0.66), while memory training’s effect on memory performance was not maintained at 10 years. Participants across the trained groups also reported less decline in everyday task function than controls. The pattern across both this trial and the Cochrane review is consistent: when cognitive training shows a durable effect, it tends to show up specifically in the domain that was trained, more reliably for reasoning and processing speed than for memory, rather than broadly transferring to overall cognitive function or to untrained abilities.

Where popular claims outrun the evidence

Marketing a brain-training app as something that meaningfully reduces personal dementia risk, rather than as a tool that may modestly improve performance on the specific trained task, goes beyond what the Cochrane review’s low-certainty, non-persistent effects and the ACTIVE trial’s domain-specific pattern actually support. In the other direction, treating the Lancet Commission’s 45% figure as a promise that addressing these factors guarantees a specific personal risk reduction misstates what a population attributable fraction, built from aggregated relative risks across a population, is designed to calculate.

Weighing this evidence against an individual situation

Dementia risk in any one person reflects a combination of age, genetics (including specific gene variants with much larger individual effects than any single modifiable factor), and the modifiable factors described here, in proportions that differ from person to person and that this population-level literature cannot specify in advance. Evaluating a specific memory concern, or deciding how a given medical history bears on these modifiable factors, is a question for a clinician who can take a full history and arrange appropriate evaluation, not something a population attributable fraction or a training-trial effect size can resolve on its own.