Risk prediction models for post-stroke cognitive impairment: a systematic review and meta-analysis

Abstract

Background The number of risk prediction models for post-stroke cognitive impairment has been increasing, but the prediction performance and clinical applicability of existing models still require further verification. Objective To systematically evaluate the published risk prediction models for cognitive impairment after stroke in patients with acute stroke. Methods The literatures on the risk prediction model of cognitive impairment after stroke in PubMed, EMBASE, web of science and Cochrane library databases on October 1, 2025 were retrieved. Extract data from the selected studies, including the collected data including author, publication year, country, time of inclusion, whether it is a multicenter study, research type, stroke type, diagnostic criteria for cognitive impairment, number of post-stroke cognitive impairment cases, model development method, validation method, missing data processing, model presentation, model prediction performance, model calibration, and predictive factors finally included in the model. The Prediction Model Risk of Bias Assessment Tool checklist was used to assess the risk of bias and applicability. Results A total of 4,715 studies were retrieved and 29 studies were included after the selection process. These studies were published from 2016 to 2023. In 29 studies, the area under the curve of development set models ranged from 0.69 to 0.97. Using random-effects meta-analysis, the descriptive summary AUC across 12 development set models was 0.85 (95% CI: 0.80–0.90), with substantial heterogeneity (I2 = 93.6%). For the 7 validation models, the descriptive summary AUC was 0.82 (95% CI: 0.76–0.87, I2 = 71.2%). Given the considerable clinical and methodological heterogeneity across studies, these pooled estimates should be interpreted as a descriptive summary of the reported AUC distribution rather than as a precise estimate of a single underlying predictive performance. Conclusion Although the descriptive summary AUC values (0.85 for development, 0.82 for validation) suggest apparently acceptable discriminative ability, all 29 included prediction model studies were judged to be at high overall risk of bias according to the PROBAST tool (particularly in the analysis domain). This likely leads to overestimation of reported model performance. Therefore, the models should be considered as preliminary tools that require further rigorous external validation before clinical application. Future research should develop a risk prediction model with larger samples and multi centers, and carry out internal and external validation. Systematic review registration https://www.crd.york.ac.uk/prospero/search, identifier: CRD42024502063.