Virtual disease landscape using mechanics-informed machine learning: Application to esophageal disorders

Area of Science:

• Gastroenterology
• Biophysics
• Computational Biology

Background:

• Esophageal disorders stem from altered mechanical properties of the esophageal wall.
• Understanding these mechanics is key to diagnosing conditions like motility disorders, eosinophilic esophagitis, reflux disease, and scleroderma esophagus.

Purpose of the Study:

• To develop a hybrid framework combining fluid mechanics and machine learning.
• To create a virtual disease landscape (VDL) for mapping esophageal mechanical behavior.
• To identify underlying physics of various esophageal disorders.

Main Methods:

• Utilized a 1D inverse model to process data from the functional lumen imaging probe (FLIP).
• Estimated mechanics-based parameters including esophageal wall stiffness, muscle contraction, and relaxation.
• Trained a neural network (variational autoencoder and side network) using these parameters.

Main Results:

• The virtual disease landscape (VDL) successfully clustered distinct esophageal disorders.
• VDL demonstrated the ability to track disease progression over time.
• The framework accurately estimated esophagogastric junction motility and mechanical work metrics.

Conclusions:

• The hybrid framework provides a novel approach to understanding esophageal mechanics.
• The VDL offers a powerful tool for distinguishing and monitoring esophageal diseases.
• Clinical applicability was shown for treatment effectiveness evaluation and patient tracking.