Machine Learning Shows Moderate Accuracy in Predicting Postoperative Nausea and Vomiting

A systematic review and meta-analysis has determined that existing machine learning models designed to predict postoperative nausea and vomiting demonstrate moderate predictive performance, proving more effective at ruling out the complication than identifying it. Published in BMC Anesthesiology by Jiaqi Wang and colleagues, the investigation systematically examined 21 studies encompassing 120 model–outcome units and a total of 285,001 patients. In the primary meta-analysis restricted to six independent validation cohorts evaluating a composite postoperative nausea and vomiting outcome, the pooled area under the receiver operating characteristic curve was 0.85, with a 95 percent confidence interval ranging from 0.70 to 0.91.

Using a bivariate random-effects model across the six independent validation cohorts, the researchers pooled diagnostic accuracy measures to evaluate model discrimination. The models achieved a pooled sensitivity of 0.74, with a 95 percent confidence interval from 0.60 to 0.85, and a pooled specificity of 0.82, with a 95 percent confidence interval from 0.69 to 0.90. The corresponding pooled diagnostic odds ratio was 12.8, with a 95 percent confidence interval between 4.6 and 35.7. Because specificity exceeded sensitivity, the authors concluded that current predictive algorithms are better at ruling out postoperative nausea and vomiting than at detecting true cases.

The study also conducted exploratory subgroup analyses across between two and nine validation cohorts, yielding pooled areas under the receiver operating characteristic curve that ranged from 0.63 to 0.91. Investigators systematically evaluated these subgroups according to differing outcome definitions, monitoring time windows, classification thresholds, geographic regions, study designs, validation methods, surgery types, and machine learning model architectures.

Methodological quality assessment using the PROBAST plus AI tool revealed widespread risks of bias across the 120 evaluated models. Exactly 50.0 percent of the models were classified as carrying a high risk of bias, 35.8 percent presented an unclear risk, and only 14.2 percent were judged to be at low risk of bias. Flaws were particularly pronounced in the statistical analysis domain, where 83.3 percent of the models exhibited a high risk of bias. A Deeks test for publication bias returned a p-value of 0.42, indicating no statistically significant publication bias, though the authors noted the statistical power of the test was limited.

Postoperative nausea and vomiting remains a frequent complication affecting postoperative recovery quality, occurring in approximately 30 percent of general surgical patients and up to 80 percent of high-risk patients. However, the systematic review observed substantial statistical heterogeneity among studies, with an I-squared of 89.2 percent for sensitivity and 94.5 percent for specificity. In addition, all included validation cohorts originated entirely from East Asia or the Middle East, with no representation from North America or Europe, a geographic restriction that limits the generalisability of the models. The authors emphasized that most current models lack independent external validation and require substantial methodological improvements before broad clinical implementation.

Subject of Research: Machine learning prediction models for postoperative nausea and vomiting

Article Title: Machine learning-based prediction models for postoperative nausea and vomiting: a systematic review and meta-analysis

Article References: Wang, J., Cao, Y., Meng, C., Li, Y., & Yin, J. (2026). Machine learning-based prediction models for postoperative nausea and vomiting: a systematic review and meta-analysis. BMC Anesthesiology. https://doi.org/10.1186/s12871-026-04407-5

Image Credits: AI Generated

DOI: 10.1186/s12871-026-04407-5

Keywords: postoperative nausea and vomiting, machine learning, prediction models, systematic review, meta-analysis, PROBAST, anesthesiology, clinical recovery, Machine, learning-based, prediction, models