Development and Validation of a Machine Learning System to Identify Reflux Events in Esophageal 24-Hour pH/Impedance...

Area of Science:

• Gastroenterology
• Medical Informatics
• Artificial Intelligence

Background:

• Esophageal 24-hour pH/impedance testing is crucial for diagnosing gastroesophageal reflux disease (GERD).
• Interpreting these studies is time-consuming and prone to variability among expert physicians.
• Currently, no machine learning tools are available for automated detection of reflux events.

Purpose of the Study:

• To develop and evaluate a novel machine learning system for automated identification of reflux events in 24-hour esophageal pH/impedance studies.
• To compare the performance of the machine learning system against existing automated software and expert physician readers.

Main Methods:

• A machine learning system was developed, incorporating signal processing and a predictive model.
• Gold-standard reflux events were established by consensus among expert physicians.
• Performance was assessed using metrics like area under the curve, sensitivity, and specificity.

Main Results:

• The machine learning system achieved an area under the curve (AUC) of 0.87, outperforming existing software (AUC 0.40) and nearing expert reader performance (AUC 0.83).
• Sensitivity for the ML system was 68.7% (vs. 61.1% software, 79.4% expert), and specificity was 80.8% (vs. 18.6% software, 87.3% expert).
• The study included 45 patients, with the model trained, validated, and tested on distinct sets.

Conclusions:

• A novel machine learning system effectively identifies reflux events in 24-hour pH/impedance studies.
• The developed system demonstrates superior performance compared to current automated software and is comparable to expert physician interpretation.
• Machine learning holds significant potential to enhance the automated analysis of esophageal pH/impedance data, improving GERD diagnosis.