Reducing False Negatives in Gastroesophageal Reflux Disease Diagnosis Through Multi-Feature Anomaly Detection of pH-Impedance Signals

Gastroesophageal reflux disease (GERD) diagnosis traditionally relies on acid exposure time (AET) obtained from 24-h multichannel intraluminal impedance-pH (MII-pH) monitoring. However, such single-metric approaches often fail to capture borderline or transient reflux events, leading to false negatives (FNs) that impede timely and appropriate treatment. To address this limitation, we propose a machine learning-based framework that integrates statistical and waveform-derived features from pH signals to enhance anomaly detection. Using one-class support vector machine and support vector data description models trained on real-world MII-pH datasets, the framework achieved an \(\:{F}_{3}\) score of approximately 0.9 and identified nearly twice as many abnormal cases as the conventional AET criterion. Explainable AI techniques, using Shapley additive explanations values, showed that features such as kurtosis and peak-to-peak amplitude contributed significantly to the identification of subtle reflux patterns. Furthermore, several FN cases that were not detected by AET or the DeMeester score were retrospectively validated through expert review and correlated clinical indicators. This approach significantly improves the detection of potential FN cases overlooked by AET-based methods, thereby contributing to more effective GERD diagnosis and treatment in clinical practice.