Development and Validation of a Machine Learning Model for Postoperative Nausea and Vomiting in Patients With Liver Cancer
Development and Validation of a Machine Learning Model for Postoperative Nausea and Vomiting in Patients With Liver Cancer
PONV management has been recommended as a necessary part of enhanced recovery protocols during the perioperative period, and PONV risk assessment is, therefore, a necessary first step in determining the number of medications or strategies for prophylaxis and treatment by considering the number of modifiable and non-modifiable risk factors. However, the external validity of two commonly-used PONV prediction models for patients undergoing liver surgery is unsatisfied, and need to be updated for liver cancer populations to better inform personalized perioperative care regime and individualized decision-making in clinical practice.
Study design This is a multi-centre, prospective cohort study performed at two tertiary teaching hospitals in Shanghai, China.
Settings and participants The study will be conducted in the liver disease department in Fudan University Zhongshan Hospital and Shanghai cancer center. Patients will be prospectively consecutively recruited between may 2024 and August 2024. Patients who are diagnosed with liver cancer and underwent hepatectomy will be eligible. Inclusion criteria are age older than 18 years, planned admissions and elective surgery, and staying at least 24 hours in the surgical unit. We exclude patients with cognitive impairment and patients who had nausea and vomiting related to other existing diseases, such as gastroesophageal reflux disease. We also exclude patients who had severe postoperative complications, including massive abdominal hemorrhage, hepatic encephalopathy and portal vein thrombosis.
Our study aims to develop PONV prediction model. The rule of thumb in logistic modelling is that there should be a minimum of 10 events per predictor variable (EPV). According to a previous study, the incidence of PONV is 48.3% [22] and there are a total of 20 predictor variables, therefore, the required minimum sample size is 414. Considering 15% of the sample loss, we would target to recruit 476 to minimize the limitation of a small number of events of PONV.
Data collection PONV risk factor assessments All enrolled patients will receive preoperative PONV risk assessments by the second author and the third author 1 day before surgery. The baseline demographic data and medical history will be recorded. The potential predictors include female sex, nonsmoking history, history of motion sickness or PONV, age, sex, history of smoking, history of motion sickness, history of PONV, duration of surgery, the use of postoperative opioids, the style of surgery and type of surgery, the numbers and time of portal vein occlusion and the use of postoperative opioids. We defined smoking history as nicotine use before surgery, history of motion sickness as nausea or vomiting when travelling in a car/boat/train/plane, and use of postoperative opioids as the use of opioids within the 24 hours after surgery. Nonsmoking history and history of motion sickness or PONV were collected by interviewing patients and family members, while the use of postoperative opioids will be determined by checking the hospital information system to review records and anaesthetic protocols.
Outcome measures Postoperative nausea and vomiting will be assessed every hour during the first two hours, every two hours for the following four hours and every four hours until the 24th hour by the first and the second authors to ensure high-quality data collection. PONV will be evaluated on a four-grade scale from 0 (no nausea and no vomiting), 1 (having nausea but no vomiting), 2 (having vomiting without stomach contents) to 3 (having vomiting with stomach contents). A patient will be considered to have had PONV if his or her PONV grade is 2 or more within the first 24 postoperative hours. PONV will be assessed by the first author, who is blinded to the results of PONV risk assessments.
Statistical analysis The statistical analysis will be performed using SPSS 25.0 and Pyhthon software version 4.0.1. Continuous variables were analysed by using descriptive statistics (median, interquartile range, or mean [SD]); categorical data were analyzed as proportions (number, percentage). According to the ratio of 8:2, the data will be divided into training sets and validation sets, and the prediction model of the random forest algorithm will be constructed, the hyperparameter optimization will be carried out, and the importance of each predictor in the random forest model will be calculated by Gini coefficient. The differentiation of the prediction model will be evaluated based on the ROC curve, the calibration curve evaluated the calibration degree of the model, and the prediction performance of the model for PONV will be evaluated. A P value of less than .05 will be considered statistically significant.
Inclusion Criteria:
Exclusion Criteria:
zhang.yx@aliyun.com13816881925