Clinical prediction rules for cognitive outcomes post-stroke : an updated systematic review and meta-analysis

Abstract

Background: Survivors of stroke are at a higher risk of cognitive syndromes, including dementia and delirium. Timely identification of those at-risk for cognitive syndromes could ensure better clinical management and implementation of risk reduction strategies. This study updates and appraises current evidence on prognostic accuracy of multicomponent risk models for post-stroke cognitive syndromes. Methods In this updated systematic review, we searched multidisciplinary electronic databases between November 2019 and October 2024 for relevant studies. An updated search was conducted on May 30, 2025. Studies were included if they described a multicomponent risk prediction tool developed in a stroke population (aged ≥18 years), free of cognitive impairment/dementia at baseline, with no exclusions on language. All study designs of primary research were eligible provided the study reported a multicomponent model at any point to predict participant cognitive outcomes i.e., incident cognitive impairment, dementia or delirium. Multicomponent refers to having more than one feature in the model e.g. if the study only reported the discriminatory accuracy of a cognitive score this was not eligible. All studies had to report sufficient discriminative performance metrics to assess model performance. Data were extracted from selected studies using a pre-specified proforma. Risk of bias was assessed using the Prediction model Risk of Bias Assessment Tool (PROBAST), certainty of evidence by GRADE, and between-study heterogeneity via I-squared (I2) statistics. Our study was preregistered with PROSPERO (CRD42024601845). Findings From 16,259 articles, 20 new studies contributed 31 models for post-stroke cognitive impairment and/or dementia and six models for post-stroke delirium with most developed in Asia (n = 12). Most models (n = 10) used logistic regression, with some using machine learning methods (n = 5). Development cohorts were small (mean n = 677). The pooled c-statistic for post-stroke cognitive impairment and delirium were 0.81 (95% CI 0.77–0.85, I2 95.7%) and 0.85 (95% CI 0.77–0.93, I2 52.7%), respectively. Three models externally validated (C-statistic: 0.72–0.91); and two models underwent temporal validation (AUC 0.81–0.82). Eight studies included measures of calibration which all demonstrated good calibration. Most studies (n = 17) were deemed to have low risk of bias and applicability concerns but overall certainty of evidence by GRADE was low. Interpretation Development of risk models to predict cognitive syndromes post-stroke has increased. Development cohorts remain small, largely developed in Asia with very few assessing model transportability. Future studies should pool data and utilise the potential of routinely collected large datasets. Stakeholder engagement and cost-effectiveness of risk-stratified interventions are needed prior to clinical implementation.