Performance of prediction models for delayed union and nonunion after fracture: a systematic review and meta-analysis.

BACKGROUND

Delayed union and nonunion are common and costly complications after fractures, yet early risk stratification remains challenging. We systematically reviewed and meta-analyzed multivariable prediction models for compromised fracture healing.

METHODS

MEDLINE, EMBASE, CINAHL, SinoMed, and CNKI were searched from inception to 30 November 2025. Studies developing or validating models predicting delayed union, nonunion, or related healing outcomes after fractures were included. Risk of bias was assessed with PROBAST and certainty of evidence with GRADE. Discrimination was summarized using the area under the receiver operating characteristic curve (AUC), pooled with random-effects models by validation tier (apparent performance, internal validation, external validation), and further described by anatomical subgroup. The protocol was registered in PROSPERO (CRD420251252244).

RESULTS

Seventy-seven studies reporting 97 model entries were included across appendicular, proximal femoral, multisite, and axial skeletal fracture settings; 56 models reported apparent performance only, 28 had internal validation, and 13 underwent external validation. Pooled AUC was 0.88 (95% CI 0.86-0.90) for apparent performance, 0.72 (95% CI 0.41-0.90) for internal validation, and 0.81 (95% CI 0.73-0.86) for external validation, with substantial heterogeneity. Apparent performance exceeded validated performance (mean optimism 0.052).

CONCLUSION

Existing models often show optimistic apparent discrimination; however, discrimination alone is insufficient to justify clinical implementation, especially given limited external validation, sparse calibration reporting, major clinical and anatomical heterogeneity, and high risk of bias. Future studies should prioritize clearer outcome definitions, robust methodology, transparent reporting, independent external validation with calibration assessment, and evaluation of clinical utility before these models are used to guide care.