Evaluating the accuracy of machine learning in predicting postoperative flap complications: A meta-analysis

Abstract

Objective: To conduct a systematic review and meta-analysis to determine the sensitivity and specificity of machine learning models in predicting complications following flap surgery.

Data sources: Five major databases, including MEDLINE, PubMed, EMBASE, EMCARE, and Google Scholar, were searched to identify relevant studies.

Review methods: We identified 49 records after removing 12 duplicates; 37 studies were screened and 32 were excluded, leaving 5 studies that were included. The total patient number was 7734 and was analysed using 10 machine learning models. For eligible studies, we extracted data on sensitivity, specificity, accuracy, and complication rates, focusing on the predictive performance of machine learning algorithms in identifying postoperative flap complications. Studies were evaluated using the QUADAS-2 tool; inclusion required reporting quantitative metrics such as sensitivity, specificity, or area under the receiver operating characteristic curve.

Results: The pooled sensitivity was 41.9% (95% CI: 41.0%-42.7%) and pooled specificity was 78.6% (95% CI: 78.2%-79.1%). Subgroup analysis showed the highest specificity in gradient boosting (GB) models (84.6%) and highest sensitivity in artificial neural network models (49.8%).

Conclusion: Machine learning models demonstrate high specificity in predicting flap failure (correctly exclude the presence of flap failure), specifically in the GB model. However, the relatively low sensitivity remains a concern. This meta-analysis was registered in the International.

Prospective register: This systematic review is registered in PROSPERO under the ID CRD42024563930.

Keywords: Flap surgery; Head and neck; Mortality or morbidity; Supervised machine learning.