How to Prevent Dementia: Risk Factors Across the Life Course, Ranked, and What People Search at Each Age

In plain words: A handful of factors hold up for dementia; many popular ones fade once small studies are discounted.1. Small studies inflate the ratios. The least precise third of studies pools to 1.64 against 1.14 for the most precise third (dementia of any kind, 422 ratios, 386 studies); the regression multiplier per unit of standard error is 5.6 (95% CI 3.5 to 8.9). Adjustment, publication year, region and ratio type do not explain it (null).2. Factors that stand. Depression 1.83 (95% CI 1.45 to 2.31), k = 25; diabetes 1.51 (95% CI 1.33 to 1.71), k = 28; hearing loss 1.25 (95% CI 1.16 to 1.34), k = 15; higher physical activity 0.81 (95% CI 0.73 to 0.89), k = 18; healthy diet 0.78 (95% CI 0.72 to 0.85), k = 11. Prediction intervals include 1 for all but a few families.3. Factors that do not stand. Obesity (1.37 (95% CI 1.23 to 1.53), k = 14) falls to 1.03 in the precise half; social isolation (1.36 (95% CI 1.10 to 1.69), k = 8) to 1.10; vision impairment, poor oral health, higher education and vaccination do not survive either.4. Life stage and lead time. Late-life exposure gives the same ratio as other stated stages within family (1.01 (95% CI 0.87 to 1.18) for all-cause dementia; null in the cognitive and any-dementia lenses and in families prone to reverse causation). The midlife rule failed its blind check, follow-up length is stated in too few rows to test, and only 14 of 422 any-dementia rows are stated as midlife.5. After diagnosis. Dementia and death pool to 1.55 (95% CI 1.25 to 1.93), k = 16; hospital admission is unsupported (interval includes 1).6. Attention. Search for dementia in 2025 to 2026 was 1.82 times its 2019 level (dementia and hearing loss 3.42; weather control 1.13) while the news rate per 100,000 articles changed by 1.13 times; the detrended news-search correlation is -0.09.7. Context. The research layer holds 26,619 core studies from 2017 to 2025, mostly of older adults (2,920 against 417 middle-aged where age is stated).8. Method. Ratio rows were re-read from quotes by a strict parser (288 of 4,672 candidates removed) and pooled within family by DerSimonian-Laird random effects with prediction intervals; every field and flag used was read blind by a second AI reader at 70% agreement or better; fields that failed are not used.Research preprint, not peer reviewed. The ratios are associations reported in the cited studies and are not medical advice. Author: Narendra Kumar (ORCID 0009-0001-7468-6100). This paper was not produced entirely by AI; the blind second readers used for validation are AI readers, named in the paper.Also deposited at: https://zenodo.org/record/24124579