Prediction Models for Postoperative Delirium After Cardiovascular Surgery: A Systematic Review and Critical Appraisal

Postoperative delirium of cardiovascular surgery (PODOCVS) is an acute brain dysfunction, which will lead to increased postoperative complications and high mortality rates. Using prediction models for early identification and intervention could improve outcomes and resource utilisation. However, the quality and applicability of existing models remain unclear, and healthcare professionals are often unsure which model should be recommended to cardiovascular surgery (CS) patients in a specific setting.

To systematically review and critically evaluate the development, performance and applicability of available prediction models for PODOCVS.

We searched multiple databases from inception to 12 August 2024, to identify multivariate predictive models for PODOCVS. Prospective or retrospective cohort studies were eligible if they created and validated delirium prediction models or scoring systems. We included studies involving adults undergoing CS, excluding studies that externally validated existing models or did not validate the models. Data extraction used the CHARMS checklist, and model quality was assessed with the PROBAST.

Of 3967 screened studies, 24 described 67 prediction models. The incidence of postoperative delirium following CS ranged between 3.6% and 36%. Logistic regression was the most commonly used modelling technique. Internal model validation was carried out in 17 studies (70.38%), with the bootstrap and random split methods being the most commonly used. The most commonly used predictors were age, EF value, extracorporeal bypass time and ICU stay duration.

While existing prediction models for PODOCVS demonstrated good discriminatory ability, they were found to have a high risk of bias according to PROBAST. All studies had a low risk of bias in predictor and outcome domains, but analysis domains were at high or unclear risk due to underreporting. The utility and generalisability of these models are uncertain due to this bias, heterogeneity in predictors and a lack of clinical application studies. Therefore, current models are not recommended for clinical use.

Early delirium prediction is crucial for timely intervention and can provide actionable information to nurses. However, the current models are not sufficiently reliable for clinical application. Future work should focus on improving methodological rigour and performing robust internal and external validation.