Explainable Machine Learning for Prediction of Early Postoperative Nausea and Vomiting After General Anesthesia

Postoperative nausea and vomiting (PONV) remains one of the most common adverse effects associated with anesthesia care. This study aimed to explore the feasibility of applying machine learning models trained exclusively on routinely available non-invasive clinical indicators to predict early PONV risk. Explainable artificial intelligence techniques were also employed to identify the most influential predictors of early PONV.

A retrospective dataset from Cathay General Hospital, including 927 patient cases and 16 non-invasive clinical indicators, was used to investigate early PONV risk prediction. This study evaluated the predictive performance of several traditional machine learning models, deep learning architectures, and ensemble learning methods to compare their classification capabilities.

Overall, the models demonstrated moderate discriminative performance. The random forest model achieved an accuracy of 83.5% with balanced precision (80.81%) and recall (83.5%), while the logistic regression model attained an AUC of 0.6905. Analysis of positive SHAP values identified the top 7 most influential predictors of early PONV. These included pharmacologic interventions (eg, neostigmine), pre-existing comorbidities (eg, history of nausea and vomiting, history of cardiovascular disease), demographic characteristics (eg, gender), postoperative pain, and anesthetic and surgical factors (eg, type of surgery and duration of anesthesia). Moreover, SHAP analysis revealed that the use of dexamethasone was negatively associated with the predicted risk in the model, suggesting its potential protective role in the prevention of early PONV.

By generating explainable outputs, this study bridges the gap between algorithmic prediction and clinical decision-making, allowing anesthesiologists to better recognize underlying risk factors and make informed, evidence-based decisions in perioperative management.