Those two problems are measurement error and reverse causation. Most previous studies have relied on people’s own reports of how long they sleep, which are notoriously imprecise. Worse, Alzheimer’s disease has a long preclinical phase, and changes in sleep patterns can be an early symptom of the disease process itself. If disturbed or shortened sleep is a consequence of brewing brain pathology rather than a cause of it, observational studies will produce misleading associations no matter how carefully they adjust for confounders. Some meta-analyses have even suggested that unusually long sleep raises dementia risk, a pattern many researchers suspect reflects prodromal disease rather than a genuine harmful effect of sleeping more.

The Oxford team’s strategy was to replace the behavioral measurement of sleep with a genetic one. Because genetic variants are fixed at conception, a polygenic score for sleep duration cannot be altered by early Alzheimer’s pathology, lifestyle changes, or anything else that happens later in life. The researchers first conducted a fresh genome-wide association study of sleep duration measured objectively by wrist-worn accelerometers among 77,770 white British participants who had worn an Axivity AX3 device for seven days between 2013 and 2015. Sleep was quantified using machine-learning algorithms based on random forests and hidden Markov models, converting the raw movement data into hours of sleep per 24-hour period. The average device-measured sleep duration in this group was eight hours and forty minutes.

The key technical innovation lay in how the genetic scores were built. Rather than relying on the handful of individual variants that reached genome-wide significance in previous studies, which together explained a mere 0.33 percent of the variation in sleep duration, the team used the LDpred2-auto algorithm to aggregate the effects of more than a million genetic variants across the genome. This Bayesian method re-estimates effect sizes for all variants using summary statistics and a linkage disequilibrium reference panel, capturing the collective contribution of thousands of tiny genetic influences. The resulting polygenic score explained about 2.06 percent of the variance in device-measured sleep duration in cross-validation, roughly six times more than the significant SNPs alone. SNP-based heritability of sleep duration was estimated at 15.5 percent, confirming that sleep is a highly polygenic trait.

Crucially, the researchers then applied this score to a completely separate group of 264,746 white British UK Biobank participants who had played no role in deriving it. Within this independent sample, people in the top fifth of the genetic score slept, on average, about 29 minutes longer per day than those in the bottom fifth, with predicted means ranging from eight hours 25 minutes to eight hours 54 minutes. The team then tracked Alzheimer’s diagnoses, identified through hospital records, death registers, and self-reported interviews, over a median of 12.5 years, during which 1,451 cases of Alzheimer’s disease occurred.

The results pointed in one clear direction. Compared with people in the middle fifth of the genetic score, those in the highest fifth, corresponding to roughly fifteen minutes more genetically predicted sleep per day, had a 21 percent lower risk of Alzheimer’s disease, with a hazard ratio of 0.79 and a 95 percent confidence interval of 0.67 to 0.94. The lower fifths of the score showed no statistically significant elevation in risk, and a sensitivity analysis restricted to participants aged 60 and over produced nearly identical results. The association also did not differ according to whether participants carried APOE ε4, the best-known genetic risk factor for late-onset Alzheimer’s disease, suggesting the sleep effect operates independently of this major susceptibility variant.

Because a polygenic score can influence disease risk through biological pathways other than sleep itself, a phenomenon known as horizontal pleiotropy, the team examined whether their score was associated with other established dementia risk factors from the 2020 Lancet Commission report. They found that a genetic predisposition to longer sleep was linked to lower odds of holding a university degree and of drinking alcohol daily, but higher odds of hypertension and reported depressive symptoms. When the main analysis was additionally adjusted for education, hypertension, alcohol frequency, and depressive symptoms simultaneously, the protective association with Alzheimer’s disease was essentially unchanged, indicating these factors did not confound the result. Still, the authors caution that some of these pleiotropic signals may hint at shared genetic architecture between sleep and dementia-related traits that deserves closer study.

A sensitivity analysis using parental history of Alzheimer’s disease and related dementias as a proxy outcome, a technique that boosts statistical power, found that people in the bottom two fifths of the sleep score had a marginally higher risk of proxy dementia, with a significant trend across the score distribution. The authors note that proxy measures capture both shared genetics and shared family environment, which may explain why the pattern differed somewhat from the main analysis, but they argue it nonetheless supports the hypothesis that sleep duration is genuinely related to Alzheimer’s disease.

The study’s findings contrast with several earlier Mendelian randomization analyses that found no link between sleep duration and Alzheimer’s risk, a discrepancy the researchers attribute largely to statistical power. Those studies used genetic instruments explaining only a tiny fraction of variation in sleep, yielding confidence intervals too wide to detect anything but enormous effects. By contrast, an earlier polygenic score study based on self-reported sleep found no association, which the team suggests may reflect the difference between objective and subjective sleep measurement, since the two are only moderately correlated genetically.

The authors are careful about what their results do and do not mean. The range of genetically predicted sleep differences in the study falls within the normal spectrum rather than representing pathological oversleeping, so the protective association for the longest-sleep group cannot be directly compared with the harmful long-sleep associations reported in observational studies. The score was also derived exclusively in people of European ancestry and will need validation in more diverse cohorts, such as the China Kadoorie Biobank, which has recently begun collecting accelerometer and genomic data. The researchers measured only average sleep across 24 hours, leaving open questions about nocturnal sleep versus daytime napping. Even so, the study offers a compelling new line of evidence that sleep duration may play a genuine role in Alzheimer’s risk, and it demonstrates how combining wearable sensors with modern genomic methods can help untangle one of the most contested questions in dementia prevention.

Subject of Research: Association between genetically predicted sleep duration and Alzheimer's disease risk in the UK Biobank

Article Title: Polygenic score for sleep duration in relation to the risk of Alzheimer’s disease: results from the UK biobank

Article References: Wong, A. T. Y., Floud, S., Reeves, G. K., Holmes, M. V., Travis, R. C., van Duijn, C. M., Doherty, A., & Smith-Byrne, K. (2026). Polygenic score for sleep duration in relation to the risk of Alzheimer’s disease: results from the UK biobank. Frontiers in Dementia, 5, Article 1843867. https://doi.org/10.3389/frdem.2026.1843867

Image Credits: AI Generated

DOI: 10.3389/frdem.2026.1843867

Keywords: Alzheimer's disease, sleep duration, polygenic score, UK Biobank, genetics, dementia, accelerometer, genome-wide association study, Mendelian randomization, epidemiology, wearable devices, LDpred2