JMIR Publications

ABSTRACT

Background:

Postoperative delirium (POD) remains a frequent and clinically consequential complication in older adults after hip fracture surgery. Although a growing number of multivariable prediction models have been reported, it remains unclear how these models perform in hip fracture populations, how often they have been tested beyond their derivation cohorts, and how methodologically sound the supporting studies are.

Objective:

We aimed to review published prediction models for POD after hip fracture surgery and to quantitatively synthesize reported discrimination across development, internal-validation, and external-validation datasets.

Methods:

We searched PubMed, Embase, Web of Science, and the Cochrane Library from inception to October 24, 2025, for studies that developed, updated, validated, or evaluated multivariable prediction models for POD in older adults undergoing hip fracture surgery. Studies of elective arthroplasty were excluded. Risk of bias was assessed with the Prediction Model Risk of Bias Assessment Tool (PROBAST). Areas under the curve (AUCs) and C-statistics were pooled separately for development, internal-validation, and external-validation data using random-effects meta-analysis on the logit scale. Exploratory subgroup analysis, meta-regression, and leave-one-out sensitivity analysis were used to examine heterogeneity and the stability of pooled estimates.

Results:

We included 24 studies, and 21 contributed development AUCs to the primary meta-analysis. The pooled development AUC was 0.833 (95% CI 0.774-0.879), with substantial heterogeneity (I2=93.2%) and a wide 95% prediction interval (0.483-0.964). Performance was lower in validation datasets, with pooled AUCs of 0.784 for internal validation and 0.764 for external validation. After exclusion of studies with very high development AUCs (≥ 0.95), the pooled development AUC decreased to 0.796. Exploratory subgroup analysis and meta-regression did not identify a robust study-level explanation for the remaining heterogeneity, and these analyses were interpreted cautiously because several categories included few studies and correlated study-level characteristics. Most studies were at high overall risk of bias, calibration reporting was limited, and external validation was uncommon. Age, preoperative cognitive impairment, functional dependence, blood loss or transfusion, and American Society of Anesthesiologists (ASA) grade were the predictors most frequently retained across final models.

Conclusions:

Prediction models for POD after hip fracture surgery show encouraging apparent discrimination, but the present evidence still warrants cautious interpretation. Validation performance was lower than derivation performance, heterogeneity remained substantial, and most studies were at high risk of bias. Future work should focus less on repeated isolated model development and more on external validation, calibration reporting, and updating of existing models. Clinical Trial: PROSPERO CRD420251167368; https://www.crd.york.ac.uk/PROSPERO/view/CRD420251167368