Development of Machine Learning Models to Predict Postoperative GERD Symptom Resolution After Laparoscopic Nissen Fundoplication

Overview

This study aims to develop machine learning models to predict postoperative gastroesophageal reflux symptom resolution after laparoscopic Nissen fundoplication using Elastic Net regression and synthetic minority oversampling technique (SMOTE).

In patients with gastroesophageal reflux disease (GERD) refractory to medication or those expected to require long-term medical treatment, anti-reflux surgery (ARS), including Nissen fundoplication, has been performed. GERD is usually diagnosed as esophageal mucosal damage or pathological esophageal acid exposure. However, about 35% of patients with gastroesophageal reflux symptoms do not exhibit abnormal findings on esophagogastroduodenoscopy (EGD) and esophageal pH monitoring. Meanwhile, about 10% of patients with typical GERD symptoms and 30-50% of those with atypical GERD symptoms do not experience symptom improvement even after undergoing ARS. Therefore, the importance of predicting symptom improvement after ARS and appropriately selecting surgical candidates has been increasingly emphasized. Though previous studies have suggested several predictors-including the length of the lower esophageal sphincter (LES), resting pressure of the LES, and bolus exposure time-to predict GERD symptom resolution after ARS, no model comprehensively integrated the results of EGD, esophageal pH monitoring, and manometry. Elastic Net regression is a machine learning method that utilizes regularized regression analysis, combining L1 (Lasso) and L2 (Ridge) penalties. This approach makes the model relatively robust against overfitting and is suitable for datasets with a small sample size, a large number of variables, and severe multicollinearity. Synthetic minority oversampling technique (SMOTE) is a method that enhances the interpretability of the minority class in a model by oversampling minority class data using the k-nearest neighbors (k-NN) algorithm. Therefore, this study aims to develop machine learning models to predict postoperative gastroesophageal reflux symptom resolution after laparoscopic Nissen fundoplication using Elastic Net regression and SMOTE. A total of 112 patients who underwent LNF between February 2017 to February 2023 will be included in this study. Preoperative and postoperative gastroesophageal symptoms, including heartburn and regurgitation, were evaluated using the GERD Health-Related Quality of Life (GERD-HRQL) questionnaire and the Korean version of the GERD questionnaire. Postoperative symptoms were assessed at 1, 3, 6, 9, and 12 months after surgery. Patients with more than a 70% improvement in symptoms at the last follow-up will be classified as the symptom resolution group. A total of 21 models will be developed to predict the resolution of heartburn, regurgitation, or atypical symptoms using the results of manometry, 24-hour esophageal pH monitoring, or both, with seven models for each symptom. All models will also incorporate the results of EGD. Elastic Net regression and the SMOTE method will be applied to oversample the minority class and develop the model. Model performance will be validated using 5-fold cross-validation. In addition to assessing model discrimination, calibration analysis will be performed to evaluate how well the predicted probabilities align with observed outcomes. The predictive performance of conventional predictors and possible predictors, including the length of LES, resting pressure of the LES, and bolus exposure time, will be compared with the model performance of the novel model.

Conditions

Interventions

Eligibility

Locations (1)

Outcomes