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

Background Postoperative nausea and vomiting (PONV) occurs in approximately 30% of general surgical patients and in up to 80% of high-risk patients, and is a common problem affecting the quality of postoperative recovery. Machine learning (ML) has been used to develop PONV risk prediction models, but their methodological quality and clinical utility have not been systematically evaluated. Methods PubMed, Embase, Cochrane Library and Web of Science were systematically searched for studies that developed ML-based PONV prediction models and reported at least one measure of predictive performance. Risk of bias of the included models was assessed with PROBAST + AI. Independent validation cohorts were used as the unit of analysis, and sensitivity and specificity were pooled using a bivariate random-effects model. Exploratory subgroup analyses were performed by outcome definition, time window, classification threshold, region, study design, validation method, surgery type and model type. Results Twenty-one studies were included (120 model–outcome units; 285,001 patients). The primary analysis included six independent validation cohorts with a composite PONV outcome; the pooled AUROC was 0.85 (95% CI 0.70–0.91), sensitivity 0.74 (0.60–0.85), specificity 0.82 (0.69–0.90), and diagnostic odds ratio 12.8 (4.6–35.7). Across exploratory subgroups (2–9 cohorts), the pooled AUROC ranged from 0.63 to 0.91. Among the 120 models, 50.0% were at high risk of bias, 35.8% at unclear risk and 14.2% at low risk, with problems in the analysis domain being the most prominent (83.3% at high risk of bias). The Deeks test did not indicate significant publication bias ( P = 0.42), but its power was limited. Conclusions The predictive performance of existing models for PONV is moderate, with a sensitivity of 0.74 and a specificity of 0.82, suggesting that the models are better at ruling out PONV than at identifying it. Heterogeneity between studies was substantial (I² = 89.2% for sensitivity and I² = 94.5% for specificity). All included cohorts were from East Asia or the Middle East, with no cohorts from Europe or North America, which limits generalisability. Most models lacked independent external validation, and methodological quality needs improvement. ​Registration PROSPERO (CRD420261439841).