Every row of the evidence table

Direction is for carrying the risk factors. The weighted combined estimate (32.0% for seven factors) is lower than the Lancet Commission's 45% for 14.

Direction is for carrying the peripheral diseases: together they were related to about one-third of global dementia burden, led by periodontal disease, chronic liver disease, hearing loss, vision loss and type 2 diabetes. Attributable fractions assume the pooled associations are causal.

A pooled incidence and prevalence estimate with no exposure compared, so no direction is recorded.

A pooled prevalence estimate with no exposure compared, so there is no direction to record.

A threshold-derivation meta-analysis: it converges on 18 Centiloids (data-driven) and 27 Centiloids (visual reads) and flags 11-26 as a range to read with caution. It compares no exposure, so no direction exists to record.

Direction is for the test's accuracy against the reference standard; an accuracy result, not a health outcome. Plasma p-tau217 detected amyloid PET positivity with 82% sensitivity and 86% specificity, comparable to CSF p-tau217; less sensitive for tau PET.

Direction is for the test's accuracy against the reference standard; an accuracy result, not a health outcome. Plasma p-tau217 was the best-performing marker (88.1% sensitivity, 88.7% specificity, AUROC 91.1%, moderate certainty), ahead of p-tau181 and p-tau231; about 90% of studies were at high risk of bias.

Direction is for the test's accuracy against the reference standard; an accuracy result, not a health outcome. Performance varied widely across tests, from 49.3% to 91.4% pooled sensitivity and 61.5% to 96.7% pooled specificity at a single cut-point, with most studies at high risk of bias.

A pooled prevalence among carers (50.8% lonely, 37.1% socially isolated) with no non-carer comparison, so no direction for the effect of caregiving can be recorded.

Direction is for BEING a carer: carers were worse off than non-carers on every domain, with the largest gaps in depression and stress.

Benefit was larger for younger, male and non-spousal carers. The largest single effect (training the person with dementia) rests on 10 observations and the authors call it promising rather than definitive.

Digital psychological interventions reduced caregiver burden and improved self-efficacy and quality of life; the effect on depressive symptoms was not statistically significant (P = .06).

Peer support reduced distress to behavioural symptoms and increased social support and quality of life immediately after the intervention, but had no effect on depression/anxiety, loneliness, burden or competence, and no outcome was sustained at follow-up.

Direction is for the test's accuracy against the reference standard; an accuracy result, not a health outcome. Pooled sensitivity 76% and specificity 83% for dementia, with 73% / 84% in primary care; the authors call it a practical screening tool while noting high or unclear risk of bias in patient selection and timing.

Advance care planning strongly increased DNR documentation (OR 14.12) but did not change whether residents died in the nursing home (OR 1.62, 95% CI 0.72-3.66). Evidence on other end-of-life outcomes was limited and inconsistent, and no study came from a non-Western setting.

A prognosis meta-analysis: it pools time to death and to nursing home admission and compares no exposure. Women's shorter survival was attributable to later age at diagnosis.

Direction is for an older-than-chronological sleep EEG brain age: each 10-year increase in the index was associated with a 39% higher risk of incident dementia, consistent across sex and age groups.

Direction is for a healthier plant-based diet: higher overall and healthful plant-based diet scores were associated with lower dementia risk, while a higher unhealthful plant-based diet score was associated with higher risk (pooled HR 1.42).

Pooled observational studies only.

Observational. Points the opposite way to the randomised MIND diet trial.

In dementia there may be small-to-moderate benefits at six months (low-certainty evidence, I2 88-96%); in MCI ginkgo probably has little or no effect (moderate certainty); in subjective cognitive complaints the effect is very uncertain. Adverse events were similar to placebo.

Observational. Points the opposite way to the randomised MIND diet trial.

Pooled observational studies only.

Dietary intake and blood levels, pooled from observational cohorts. The much larger supplement estimate (HR 0.36) comes from one cohort's self-selected supplement users, not from a trial.

Heterogeneity was extreme (I-squared 97%), so the pooled figure hides large differences between the 10 studies.

Heterogeneity was high (I-squared 95%). Reaches a firmer conclusion than the earlier BMJ meta-analysis of largely the same literature.

The overall PM2.5 estimate includes no effect (HR 1.04, 95% CI 0.99-1.09). Only the studies that actively assessed everyone for dementia showed a clear effect (HR 1.42, 95% CI 1.00-2.02). Most studies were at high risk of bias.

High overall and leisure-time activity were associated with lower dementia risk, but occupational and commuting activity were associated with higher risk. Certainty of evidence was very low to low.

The authors call the association weak (RR 0.97), and it was absent beyond 10 years of follow-up.

Pooled across 29 cohorts. The association weakened to HR 0.85 after maximal confounder adjustment and was not significant for low-intensity activity.

The authors state the association held in longer follow-ups, which is their answer to the reverse-causation objection raised by the Whitehall II cohort.

Pools 27 small trials in people who already have dementia, and reports a much larger effect (SMD 0.69) than the single largest such trial, DAPA, which found none.

Seven small trials; working memory did not improve (P=0.14).

Direction is for a high polygenic score (excluding APOE): people in the top decile had a twofold risk of Braak stage >4 tangles and of moderate-to-severe amyloid plaque pathology at death. 91 loci identified, 16 new; 18 further loci await validation.

Locus-discovery genome-wide association study: it reports which genetic regions reached significance and no effect estimate for any exposure, so no direction is recorded. The study reports 14 new loci for clinically diagnosed AD and two from AD-by-proxy.

Direction is for carrying a higher genetic risk score: risk of AD rose 1.6- to 1.9-fold from the lowest to the highest decile, over and above age and APOE4. The main output is 75 risk loci, 42 new.

Direction is for having the impairment. Dual sensory impairment carried the highest risk (HR 1.63 for all-cause dementia, 2.55 for Alzheimer's).

Mostly observational (25 of 31 studies), so people who take up hearing aids may differ from those who do not. The authors call for randomised trials — and ACHIEVE, the trial that followed, was null overall.

Against untreated cataract. People who had surgery ended up at about the same risk as controls who never had cataracts (HR 0.84, 95% CI 0.66-1.06), and there was no short-term gain in those already cognitively impaired.

Direction is for having a vision impairment. Single vision impairment raised all-cause dementia risk but not Alzheimer's specifically; having both hearing and vision impairment carried the highest risk.

Pooled observational studies, so open to the confounding the natural-experiment studies were designed to avoid.

A statistically detectable but very small cognitive effect (ADAS-Cog SMD -0.11) alongside 107 more cases of ARIA-E per 1,000. The reviewers' own conclusion is that removing amyloid does not seem to produce clinically meaningful effects — the sharpest disagreement in this table with how the individual trials were reported.

Benefits were statistically present but every one fell below the minimum clinically important difference, against a number needed to harm of 9 for ARIA-oedema. Agrees with the Cochrane review and not with the trial reports.

Antihypertensive treatment improved cognitive scores by a small amount (SMD 0.06, moderate certainty) but the reduction in dementia incidence was not statistically significant (OR 0.89, low certainty). The drug-class rankings rest on one trial per class and are described as highly uncertain.

Antihypertensive use tracked lower Alzheimer's risk but not lower non-Alzheimer dementia risk, where treated and untreated hypertension looked the same. A single baseline blood pressure reading predicted neither.

Individual participant data from five double-blind placebo-controlled trials, which is the strongest design in this table for this exposure.

Consistent across 14 trials (I-squared 0%) but small in absolute terms: 0.39 percentage points over about 4 years.

Heterogeneity was near-total (I-squared 98.9%) and the estimate went to null in people newly diagnosed with diabetes (HR 1.01), which is where prevalent-user bias would be smallest.