Development of a diagnostic model for esophageal achalasia assessed by esophageal high-resolution manometry using artificial intelligence.

Esophageal high-resolution manometry (HRM) is an important tool for diagnosing and assessing esophageal achalasia. Artificial intelligence (AI)-assisted HRM image processing has the potential to aid in the diagnosis of esophageal achalasia. However, addressing the challenges associated with the 'black-box' problem is important. In the present study, an automated system that utilizes AI with class-activation maps to highlight diagnostic areas in HRM images was developed. A total of 211 HRM images, which led to the diagnosis of controls and patients with achalasia, were used to train the system using Resnet34, a convolutional neural network model. The diagnoses included normal, type I achalasia, type II achalasia, type III achalasia and hypercontractile esophagus based on the Chicago classification v3.0 for esophageal motility disorders. A gradient class activation map (Grad-CAM) technique was used. The discrimination model for the control and achalasia groups yielded a 100% correct response rate for evaluating the validation images (n=30). Grad-CAM analysis revealed that the model focused on the area around the lower esophageal sphincter pressure in type 1 achalasia for differentiation, closely aligning with expert perspectives. An AI-based HRM imaging assistance system may not only support physicians in distinguishing esophageal motility disorders with improved diagnostic accuracy but also serve as a novel tool that provides deeper clinical insights and highlights key interpretative features in HRM evaluations. Further large-scale validation is required to confirm its clinical utility.