Genetic versus Modifiable Risk: Predicting Dementia, Cognition, and Brain Structure in the UK Biobank

IntroductionBoth genetic and modifiable lifestyle factors contribute to dementia risk, yet their separate and joint predictive value for dementia and its intermediate markers (cognitive function and brain structure), is unclear. This study evaluated APOE {varepsilon}4 status, AD polygenic risk (excluding APOE), and the updated LIBRA2 index (14 modifiable factors) in predicting dementia, cognition, and brain imaging biomarkers. MethodsTo examine associations of APOE {varepsilon}4, PRS-APOE, and LIBRA2 with incident cases of dementia and AD, Cox models were applied to data of 345,785 UK Biobank participants at baseline (median follow-up: 13.8 years). Models were adjusted for age, sex, and other key covariates. Linear regression was used to examine associations with cognition (N=42,120) and dementia-related MRI markers (N = 31,180, such as fractional anisotropy/mean diffusivity, white matter hyperintensities, gray matter volume, hippocampus, and entorhinal cortex) among individuals without dementia. ResultsDuring follow-up, 6,300 participants developed dementia (132 cases per 100,000 person-years), including 2,803 cases of Alzheimers disease. All three predictors were independently associated with incident dementia and AD. Compared to non-carriers, APOE {varepsilon}4 carriers had a threefold increased risk of dementia (HR = 3.02, 95% CI: 2.88-3.18) and over a fourfold increased risk of AD (HR = 4.33, 95% CI: 4.01-4.68). Each 1 standard deviation (SD) increase in PRS-APOE was associated with a 26% higher risk of dementia (HR = 1.26, 95% CI: 1.23-1.29) and a 35% higher risk of AD (HR = 1.35, 95% CI: 1.30-1.40). Each 1-SD increase in LIBRA2 score was linked to a 33% increase in dementia risk (HR = 1.33, 95% CI: 1.30-1.36) and 20% for AD (HR = 1.20, 95% CI: 1.15-1.24). Beyond individual effects, we evaluated model-level discrimination using the C-index. Representative results for dementia showed modest improvement over the base model (C-index = 0.811) with PRS-APOE (0.816) or LIBRA2 (0.819), and greater improvement with APOE {varepsilon}4 (0.836). Combining APOE {varepsilon}4 and LIBRA2 yielded near-optimal discrimination (0.843), close to the full model (0.846). While APOE {varepsilon}4 was the dominant predictor of dementia diagnosis, LIBRA2 showed unique strongest predictive power for processing speed and neuroimaging markers. LIBRA2 was linked to a wide range of neuroimaging markers (e.g., WMH, WM integrity, grey matter, ventricles), while PRS-APOE was only associated with ventricular volume. APOE {varepsilon}4 was related to couple of white matter measures. Significant LIBRA2-APOE interactions were observed in relation to both all-cause dementia and AD. The association between LIBRA2 and dementia risk was stronger among APOE {varepsilon}4 non-carriers (interaction HR = 0.845, 95% CI: 0.807-0.884 for all-cause dementia; HR = 0.851, 95% CI: 0.792-0.914 for AD; both p adj < 0.001), suggesting that unfavorable lifestyle may have a greater impact in individuals without genetic risk. No significant interactions were found in cognitive or imaging models. ConclusionThe APOE {varepsilon}4 genotype was the strongest single predictor of dementia and cognitive outcomes. In contrast, the LIBRA2 index was most strongly associated with processing speed and neuroimaging markers, demonstrating its critical role in non-diagnostic brain health outcomes. These findings highlight the complementary and domain-specific contributions of genetic and lifestyle factors in dementia risk modeling.