Prediction models for postoperative cognitive dysfunction in adults: a systematic review of methodological quality...

Postoperative cognitive dysfunction (POCD) is associated with adverse outcomes. Although numerous prediction models have been developed, their methodological quality and clinical applicability remain uncertain.

We systematically searched the MEDLINE, EMBASE, and Cochrane Central Register of Controlled Trials databases for studies that developed POCD prediction models in adults. Data extraction and quality assessment were performed using CHARMS (Checklist for Critical Appraisal and Data Extraction for Systematic Reviews of Prediction Modeling Studies) and PROBAST (Prediction model Risk Of Bias Assessment Tool).

From 2,060 initial records, 13 studies comprising 14 prediction models were included. Models were developed using logistic regression (n = 8) or machine learning (n = 6). Study sample size ranged from 82 to 687. Reported discriminatory performance was high [area under the receiver operating characteristic curve (AUC) range, 0.710-0.973]; however, these results were primarily derived from internally validated or unvalidated (rather than externally validated) models. The most common predictors of POCD were age (10 models) and preoperative hemoglobin concentration (four models). Twenty-three of the 30 predictors were measured preoperatively. The primary reasons for a high risk of bias were insufficient sample size and a lack of external validation. Only two models underwent temporal external validation, yielding diminished AUC (0.617 and 0.694). Furthermore, the incorporation of non-routine predictors such as functional magnetic resonance imaging and electroencephalography compromised the generalizability of the models to routine clinical settings, particularly because the routine use of these techniques would increase both preoperative waiting times and overall medical costs. Nonetheless, if sufficient predictive power is demonstrated, these non-routine predictors may be used more routinely in the clinic.

Current prediction models for POCD show promising discriminatory performance, but their evaluation has been limited by short follow-up periods, a lack of external validation, and limited sample sizes, thus precluding their current clinical use. Future research should leverage large prospective cohorts, incorporate longitudinal multi-timepoint predictors that are routinely evaluated in clinical settings, extend follow-up to capture the full POCD course, and prioritize rigorous external validation to develop clinically applicable models.

https://www.crd.york.ac.uk/PROSPERO/view/CRD42024622477.