Risk prediction models for cognitive impairment in patients with cerebral small vessel disease: a systematic review...

This study systematically evaluates the risk prediction models for cognitive impairment in patients with cerebral small vessel disease (CSVD) and explores the predictive factors for cognitive impairment to provide effective guidance for the future development of higher-quality prediction models.

A computer-based search was conducted using the following databases: Wanfang Database, China National Knowledge Infrastructure (CNKI), VIP Database, China Biomedical Literature Database, EMBASE, Web of Science, PubMed, and The Cochrane Library. The search aimed to identify studies on risk prediction models for cognitive impairment in patients with CSVD, covering the period from the inception of each database up to 15 June 2025. A meta-analysis of the predictive factors and the area under the receiver operating characteristic curve (AUC) values of the models was performed using RevMan 5.4 and R software, respectively. The Prediction model Risk of Bias ASsessment Tool (PROBAST) was used for screening, data extraction, and assessment of the risk of bias in the included studies.

A total of 19 studies were selected for inclusion, resulting in the development of 27 risk prediction models for cognitive impairment. The AUC of all models was greater than 0.7. PROBAST assessment results indicated a high risk of bias across the studies, but the applicability of the models was relatively good. Statistical analysis using R software revealed an AUC of 0.87 (95% CI: 0.79-0.92) and 0.85 (95% CI: 0.82-0.88) for the models, indicating good predictive performance. Meta-analysis results showed that hypertension, homocysteine (Hcy), high CSVD burden, age, diabetes, and the triglyceride-glucose (TyG) index (all with p < 0.05) were the major predictors of cognitive impairment.

The performance and quality of existing risk prediction models for cognitive impairment in patients with cerebral small vessel disease (CSVD) still require improvement. The majority of the models lack external validation and appropriate calibration methods, and many are retrospective studies, which increases the overall risk of bias. Future research should focus on exploring more advanced machine learning algorithms, optimizing study designs, and emphasizing external validation to enhance the generalizability of the models. This would help build more universally applicable prediction models, thereby guiding the clinical implementation of targeted preventive measures.

https://www.crd.york.ac.uk/prospero/, identifier CRD420251074647.