Machine learning models demonstrated that the relative importance of cardiovascular and genetic predictors for long-term dementia risk differs between men and women.
Sex-specific machine learning models predict 10-year dementia risk using cardiovascular and lifestyle factors, according to study findings published in Neurology Open Access.
Cardiovascular and lifestyle factors contribute to dementia risk, but their associations with dementia may differ by sex. Researchers evaluated whether sex-specific machine learning models could predict 10-year all-cause dementia risk and identify differences in the relative importance of individual risk factors.
The analysis included 43,973 dementia-free adults from 8 longitudinal cohorts participating in the Dementia Risk Prediction Project. Participants had at least 10 years of follow-up and available dementia and mortality data. The researchers divided the participants into development (n=35,176) and validation (n=8797) samples using an 80:20 split within each cohort and sex group.
The researchers developed separate Extreme Gradient Boosting (XGBoost) models for women and men. They evaluated a model containing clinical cardiovascular factors, then added lifestyle characteristics and apolipoprotein E (APOE) genotype. The researchers used Shapley additive explanations (SHAP) to assess predictor importance and evaluated model performance using the area under the receiver operating characteristic curve (AUROC) and calibration slopes.
[B]y using the machine learning prediction models, clinicians can make better informed decisions regarding prevention for dementia based on more consistent predicted risk of dementia.
Women comprised 51.5% of the development sample and 52.4% of the validation sample. The mean age was 61.8 years among women and 61.5 years among men in the development sample. Coronary heart disease was more prevalent among men than women (33.9% vs 26.6%), as were current or former smoking (64.9% vs 46.3%) and excessive alcohol use (13.6% vs 3.5%). Antihypertensive medication use was more common among women than men (37.7% vs 32.6%).
Within 10 years, 1020 women and 735 men in the development sample developed dementia. Estimated 10-year dementia risk was 7.7% (95% CI, 7.3-8.2) among women and 6.2% (95% CI, 5.8-6.6) among men. In the validation sample, dementia occurred in 239 women and 183 men, corresponding to risks of 7.3% (95% CI, 6.6-8.2) and 6.2% (95% CI, 5.5-7.0), respectively.
In the validation sample, the clinical-factor model yielded an AUROC of 0.901 (95% CI, 0.885-0.916) among women and 0.871 (95% CI, 0.845-0.897) among men. Adding lifestyle characteristics produced AUROCs of 0.901 (95% CI, 0.886-0.916) and 0.867 (95% CI, 0.839-0.895), respectively. With APOE genotype included, AUROC reached 0.906 (95% CI, 0.891-0.921) among women and 0.874 (95% CI, 0.848-0.900) among men. Calibration slopes for these full models were 1.02 (95% CI, 0.91-1.13) and 0.99 (95% CI, 0.88-1.10), respectively. Overall, lifestyle characteristics and APOE genotype provided little incremental improvement over clinical cardiovascular factors alone.
Age was the leading predictor of dementia in both sexes. Other prominent predictors were antihypertensive medication use, APOE genotype, HbA1c, stroke, and systolic blood pressure. Further, SHAP analyses revealed sex-related differences in predictor importance. For example, APOE genotype contributed more strongly to predicted dementia risk among women, whereas stroke was more important among men.
Sex differences were also apparent in the associations between predicted dementia risk and APOE genotype and systolic blood pressure. Women with 2 copies of the APOE ε4 allele had significantly higher predicted risk than men with the same genotype. Greater predicted dementia risk was associated with a systolic blood pressure threshold of approximately 130 mm Hg among women, compared with approximately 150 mm Hg among men. The researchers reported no significant differences in discrimination between the sex-specific models.
In sensitivity analyses restricted to participants older than 65 years, discrimination was lower than in the overall validation sample. With clinical and lifestyle factors and APOE genotype included, AUROC was 0.762 (95% CI, 0.730-0.795) among women and 0.762 (95% CI, 0.719-0.796) among men.
The researchers also compared XGBoost with logistic regression. Both approaches demonstrated similar discrimination, but XGBoost showed better calibration, potentially offering an advantage for estimating absolute dementia risk.
Study limitations include the predominantly Western study population, the inability to establish causal relationships from the identified associations, unavailable baseline cognitive measures across cohorts, and the assessment of all-cause dementia rather than individual dementia subtypes.
Disclosures: This research was supported by theNational Institutes of Health, including support from the National Institute of Neurological Disorders and Stroke. One study author declared affiliations with biotech, pharmaceutical, and/or device companies. Please see the original reference for a full list of disclosures.
Yu J, Stephen J, Petito LC, et al.Sex-specific machine learning–enabled risk prediction for dementia using cardiovascular risk factors.Neurol Open Access.2026;2(3):e000128. doi:10.1212/WN9.0000000000000128
