Prediction models for post-stroke delirium: a systematic review with an exploratory meta-analysis of predictors

To systematically identify and synthesize predictors of post-stroke delirium (PSD) derived from existing prediction models, and to assess the methodological quality of these studies using PROBAST.

A comprehensive systematic search was conducted in nine databases from inception to April 2026. Studies developing or validating prediction models for PSD were included. Data extraction was guided by the CHARMS checklist. Methodological quality and risk of bias were assessed using the Prediction Model Risk of Bias Assessment Tool (PROBAST). Meta-analysis was performed to pool the effect sizes of predictors and the area under the receiver operating characteristic curve (AUC).

Sixteen studies (24 models) with sample sizes ranging from 100 to 14,475 were included. Model discrimination was moderate to good, with reported AUC values ranged from 0.72 to 0.94. The meta-analytic pooled AUC was 0.83 (95% Confidence interval: 0.81-0.85). Age, NIHSS (National Institutes of Health Stroke Scale score), neutrophil-to-lymphocyte ratio, visual impairment, and infection were identified as common significant predictors. PROBAST assessment revealed a high overall risk of bias in all studies, primarily due to methodological shortcomings in the analysis domain. Calibration was assessed in six studies with acceptable performance, whereas clinical utility was rarely evaluated.

This study highlights several important predictors of PSD. However, due to the high risk of bias, the reliability of existing models remains uncertain. Although the pooled AUC of 0.84 suggests moderate to good discrimination, its performance in individual clinical settings may vary markedly. Future studies should adhere to unified PSD diagnosis criteria, employ robust validation strategies, and explore advanced modeling techniques to improve model reliability and clinical utility.