A Machine Learning Prediction Model for Delayed CIPONV

Overview

Postoperative nausea and vomiting (PONV) can lead to serious postoperative complications, but most symptoms are mild. Clinically important PONV (CIPONV) refers to PONV symptoms that have a significant impact on the patient's well-being and recovery. Present predictive systems for PONV are mainly concentrated on early PONV. However, there is currently no suitable prediction model for delayed PONV, particularly delayed CI-PONV. This study aims to develop and validate a prediction model for delayed CI-PONV using machine learning algorithms utilizing perioperative data from patients undergoing laparoscopic gastrointestinal surgery. All 1154 patients in the FDP-PONV trial will be enrolled in this study. Delayed CIPONV is defined as experiencing CIPONV between 25-120 hours after surgery. After selecting the modeling variables from 81 perioperative clinical features, six machine learning models are established to generate the risk prediction models for delayed CIPONV. The area under the receiver operating characteristic curve, accuracy, sensitivity, specificity, positive predictive value, negative predictive value, F1 score and Brier score are used to evaluate the model performance. Shape Additive explanation analysis was conducted to evaluate feature importance.

The website https://mvansmeden.shinyapps.io/BeyondEPV/ was used for sample size calculation, considering 6 candidate predictors, an event fraction of 0.14, and a criterion value for reduced mean predictive squared error of 0.03. The calculated sample size is 1080, with a minimally required expected event per variable of 25.1. Therefore, a sample size of 1154 patients is deemed sufficient to support the inclusion of 6 predictors in the development of the predictive model. A total of 81 variables, including demographics, comorbidities, laboratory findings, as well as information related to anesthesia and surgery, are prospectively collected in the FDP-PONV trial and considered as potential predictive factors in this study. The least absolute shrinkage and selection operator method is used to identify clinically significant variables. Further selection of the final predictors is performed using stepwise regression based on the Akaike Information Criterion. The entire dataset is randomly divided into a training set and a validation set in a ratio of 7:3. Six machine learning models, namely logistic regression, random, extreme gradient boosting, k-nearest neighbor, gradient boosting decision, and multi-layer perceptron, were developed to create risk prediction models for delayed CIPONV. The performance of the models is assessed by comparing the area under the receiver operating characteristic curve, accuracy, sensitivity, specificity, positive predictive value, negative predictive value, F1 score, Brier score and calibration curve. Bootstrap resamples is conducted 1000 times on the training cohort to evaluate the predictive model's performance. Decision curve analysis is conducted to assess the clinical applicability of the model. The SHapley Additive Explanations library (SHAP) is used to interpret the prediction model.