Yueliang Sun, Jiankang Wang, Fengling Sun, Mingpeng Shi, Xinxin Tian, Hang Shang, Wenxuan Wang, Tianying Chang, Jiajun Zhao, Yongsheng Huang, Yingzi Cui, Jiajuan Guo
Seventeen studies reporting 25 models were included. Most models were developed in China using cross-sectional data, with cognitive outcomes assessed using varied instruments. Reported AUCs ranged from 0.344 to 0.938; most multivariable models exceeded 0.70. Internal validation was reported in 15 studies, whereas only four conducted external validation and 11 assessed calibration. Sixteen studies were judged at high risk of bias. Reproducibility was limited: four studies provided sufficient information for independent calculation, eight were partially reproducible, and five were insufficiently reported. Common predictors included older age, lower education, depression, obesity, longer hypertension duration, greater hypertension severity, and adverse metabolic indicators.
BACKGROUND: Hypertension is strongly associated with cognitive impairment. Existing tools include models designed to identify prevalent cognitive impairment and a smaller number developed to predict future risk. However, their methodological quality, validation status, and reproducibility have not been systematically assessed. We conducted a systematic review and critical appraisal of existing models designed to identify cognitive impairment in patients with hypertension.
METHODS: Chinese and English databases were searched from inception to December 23, 2025. Studies that developed, updated, or validated a multivariable model, nomogram, score, or machine-learning tool for cognitive impairment, mild cognitive impairment, cognitive decline, subjective cognitive decline, or dementia among adults with hypertension were included. Data were extracted using CHARMS, and risk of bias was assessed with PROBAST. The synthesis emphasized model purpose, temporality, outcome definition, discrimination, calibration, validation, risk of bias, and reproducibility. Reported c-statistics were summarized descriptively because of substantial heterogeneity in study design, outcome definitions, cognitive assessment tools, and repeated model estimates from the same datasets.
RESULTS: Seventeen studies reporting 25 models were included. Most models were developed in China using cross-sectional data, with cognitive outcomes assessed using varied instruments. Reported AUCs ranged from 0.344 to 0.938; most multivariable models exceeded 0.70. Internal validation was reported in 15 studies, whereas only four conducted external validation and 11 assessed calibration. Sixteen studies were judged at high risk of bias. Reproducibility was limited: four studies provided sufficient information for independent calculation, eight were partially reproducible, and five were insufficiently reported. Common predictors included older age, lower education, depression, obesity, longer hypertension duration, greater hypertension severity, and adverse metabolic indicators.
DISCUSSION: Existing models show some ability to distinguish hypertensive patients with and without cognitive impairment, while evidence supporting future-risk prediction remains limited. Future work should prioritize prospective cohorts, standardized outcome definitions, external validation, calibration, and transparent reporting of equations, intercepts, scoring rules, code, and model objects.
SYSTEMATIC REVIEW REGISTRATION: https://www.crd.york.ac.uk/PROSPERO/view/CRD420251157136, identifier CRD420251157136.