A cascade deep learning model for diagnosing pharyngeal acid reflux episodes using hypopharyngeal multichannel intraluminal Impedance-pH signals

Abstract

Detecting pharyngeal acid reflux (PAR) episodes from 24-h ambulatory hypopharyngeal multichannel intraluminal impedance-pH (HMII-pH) signals is crucial for diagnosing laryngopharyngeal reflux (LPR). Currently, a lack of effective software for PAR episode detection requires time-consuming manual interpretation, which is prone to inter-rater variability. This study introduces a deep learning-based artificial intelligence (AI) system for PAR episode detection and diagnosis using HMII-pH signals. Ninety patients with suspected LPR and 28 healthy volunteers underwent HMII-pH testing in three Taiwanese medical centers. Candidate PAR episodes were defined as esophagopharyngeal pH drops exceeding 2 units, with nadir pH below 5 within 30 seconds during esophageal acidification. A consensus review by three experts validated 84 PAR episodes in 17 subjects. Data preprocessing identified 225 candidate PAR episodes, including 84 PAR episodes and 141 swallows/artifacts, were divided into training, validation, and test datasets (6:2:2 ratio). Three cascade deep learning AI models were trained. Among them, the cascade Multivariate Long Short-Term Memory with Fully Convolutional Network (MLSTM-FCN) model performed best in the test dataset. At the episode level, this model achieved 0.936 accuracy, 0.941 precision, 0.889 recall, 0.966 specificity, 0.914 F₁ score, and 0.864 Matthew's correlation coefficient (MCC). For subject-level evaluation, the corresponding metrics were 0.917 accuracy, 1.000 precision, 0.818 recall, 1.000 specificity, 0.900 F₁ score, and 0.842 MCC. In conclusion, the cascade MLSTM-FCN model demonstrates robust accuracy in diagnosing PAR episodes from HMII-pH signals, offering a promising tool for efficient and consistent PAR episode detection in LPR diagnosis.