(0808) Development of a Machine Learning Model to Predict Therapeutic Responses in Laryngopharyngeal Reflux Disease

Introduction: Laryngopharyngeal reflux (LPR) is a common disease, but the response to treatment varies from patient to patient. Thus, it is important to find key factors influencing treatment response. This study leverages various machine learning models to identify key factors and predict treatment response in LPR.

Methods: Patients who visited one tertiary medical center with typical LPR symptoms, and underwent laryngoscopy and 24-hour multichannel intraluminal impedance–pH (MII-pH) monitoring were investigated prospectively between 2014 and 2023. Patients who were prescribed high-dose proton pump inhibitors (PPIs) twice daily and followed up for at least 2 months were enrolled. Patients showing at least a 50% decrease in the reflux symptom index (RSI) score after treatment were defined as responders. The dataset comprised patient medical history, treatment periods, baseline and follow up RSI, laryngoscopic findings, and various parameters in 24-hour MII-pH monitoring. We utilized a variety of machine learning models, including logistic regression, random forest, support vector machine (SVM), and gradient boosting.

Results: A total of 157 patients were enrolled in this study, and divided into responders (n = 104) and non-responders (n = 53). The accuracy and area under the receiver operating characteristic curve were highest in the logistic regression model, followed by the random forest model, showing their effectiveness in predicting treatment outcomes in patients with LPR. The feature importance analysis revealed various factors such as proximal recumbent reflux time and proximal recumbent weak acid time, influencing treatment response in LPR.

Conclusions: This study shows that usefulness of machine learning models in predicting treatment responses in LPR. The results highlight the importance of a data-driven approach in clinical decision-making, particularly in enhancing therapeutic outcomes for LPR patients. Our findings suggest that integrating machine learning models into clinical practice could significantly improve the management of LPR.