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

Comprehensive Summary

Researchers analyzed a retrospective dataset of 927 patients from Cathay General Hospital. They tested 12 different models, including traditional Logistic Regression to complex Deep Learning and Ensemble methods (Random Forest). Of these models, the Random Forest achieved the highest overall accuracy at 83.5%. This model also had the best discriminative capacity with an AUC of 0.6905. The researchers used SHAP (Shapley Additive exPlanations) values to find out exactly what factors pushed a specific patient’s risk score up or down. The main risk predictor they found was neostigmine use, which is commonly used to reverse muscle relaxers during surgery. Other dominant risk factors were prior history of nausea/vomiting, female gender and younger age, and surgery lasting longer than 100 minutes and longer than 120 minutes of anesthesia. They also discovered that dexamethasone, as identified by the AI, provided a significant weight against the development of PONV (post-operative nausea and vomiting). It is important to note that the researchers did acknowledge that the models used only demonstrated moderate performance, and the current indications do not point directly to immediate clinical application. Instead, future work should look for prospective external validation to properly indicate real-world clinical impact.

Outcomes and Implications

This study indicates a clear pharmacological change from neostigmine to sugammadex. For high-risk patients identified by models like this one, choosing sugammadex for neuromuscular blockade reversal could significantly improve the PACU experience by decreasing N/V. This study highlights the importance of including standard risk scores, instead of a one-size-fits-all approach, using AI models to provide real-time risk alerts at the end of a case, which could prompt providers to add a late-stage antiemetic if the math changes peri-operatively. This could also induce new pre-operative patient discussion protocols relating to risk, letting the patient know that based on their history and length of surgery, the care team estimates a higher risk score, and will act accordingly. An important discovery also was that explainable AI like this one, can align with clinical institution standards such as confirming the efficacy of Dexamethasone, which will help increase provider trust and widespread adoption.