Robust Methods to Detect Abnormal Initiation in the Gastric Slow Wave from Cutaneous Recordings

Area of Science:

• Gastroenterology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Upper gastrointestinal (GI) disorders, including gastroparesis (GP) and functional dyspepsia (FD), are common, affecting 3% and 10% of the US population.
• Gastric slow wave abnormalities disrupt food propulsion, complicating diagnosis and treatment of GP and FD.
• Traditional spectral analyses struggle to identify these abnormalities due to their frequency overlapping normal patterns.

Purpose of the Study:

• To develop a robust method for classifying normal versus abnormal gastric slow wave initiation location and propagation patterns.
• To improve the identification of gastric motility disorders using advanced computational techniques.
• To enhance diagnostic accuracy for upper GI disorders by addressing limitations in current analysis methods.

Main Methods:

• Utilized a 3D convolutional neural network (CNN) trained on multi-electrode cutaneous recordings.
• Employed transfer learning to build a model robust to variations in abnormality location and recording start times.
• Implemented a progressive training approach, starting with simpler models and increasing complexity.

Main Results:

• Achieved an average classification accuracy of 80% for identifying abnormal slow wave patterns.
• Demonstrated Type-I error probabilities below 5% across various spatial abnormality locations.
• Successfully classified normal versus abnormal slow wave patterns in silico, even with challenging spatial variations.

Conclusions:

• The developed transfer learning-based 3D CNN method effectively classifies gastric slow wave abnormalities.
• This AI-driven approach offers a promising solution for the challenging diagnosis of upper GI motility disorders.
• The method shows robustness and high accuracy, paving the way for improved clinical diagnostic tools.