Machine Learning Algorithm Applied to pH-Impedance Monitoring Tracings Identifies Pathologic Supragastric Belching

Supragastric belches (SGB) are not detected by commercial software and are cumbersome to identify manually. We determined whether machine learning can accurately identify SGBs.

pH-impedance studies from a convenience cohort (20 patients each of pathologic GERD, excessive SGB, and healthy volunteers) were randomly assigned to training (2), validation (4), and testing (4) from each group. First, studies were manually reviewed to mark SGBs, then converted to .csv files and input into a convolutional neural network (CNN) developed de novo using open-source software. The training set was utilized for hyperparameter and epoch optimization. Subsequently, clinical utility was assessed using a validation set followed by test set to confirm accuracy. False positives were reviewed, allowing recognition of novel CNN-detected events.

339 SGB events (median 8.0) were manually identified in the GERD cohort, > 2000 in the SGB cohort and 163 (median 3.0) in healthy volunteers. When analyzed by group, sensitivity was highest in healthy volunteers and lowest in the GERD cohort, while false positive rates were similar. Unique new SGB events were most common in the SGB cohort (p < 0.001 compared to the other cohorts). On ROC analysis, machine learning segregated SGB patients from GERD and healthy volunteers with high accuracy (AUC 0.859, p = 0.001) at an optimal threshold of 11 events, with sensitivity of 100% and specificity of 62.5%.

In this proof-of-concept investigation, a machine learning algorithm accurately identifies SGB events, including events missed by manual review, and reliably segregates SGB patients from healthy volunteers and GERD patients.