Advancing GI motility with AI: A transatlantic multicentric study of cross-site and cross-device manometry analysis

Abstract

Since its early development, manometry has been fundamental in the diagnostic workup of functional esophageal and anorectal motility disorders. Nevertheless, these exams are hindered by complex data analysis and high interobserver variability. Additionally, there is still a significant gap between low (where conventional manometry is still performed) and high-resource settings (where high resolution devices prevail). In this context, artificial intelligence offers the promise of creating an unified framework for interoperable manometry classification. This study aims to develop and validate artificial intelligence models for classification of motility disorders in both conventional anorectal manometry (CARM), high-resolution anorectal manometry (HR-ARM) (according to London Classification) and high-resolution esophageal manometry (HREM) (according to Chicago Classification). Methods: Multiple machine learning (ML) models were developed and tested after inclusion of exams from 4 centers in 3 countries. 827 ARM exams were used for development and testing of 4 ML models, with 90% of the data used for training and 10% for testing. 960 HR-ARM exams were used for development of ML models according to London Classification, with a 80%/20% patient split. Finally, 618 HREM exams were used for development of ML models according to Chicago Classification, with a 80%/20% patient split. Model’s performance was evaluated through metrics as accuracy and the area under the receiver-operating characteristic curve (AUC-ROC). Results: The xGBClassifier discriminated obstructed defecation from fecal incontinence in ARM with an overall accuracy of 84.6% and AUC-ROC of 0.939. Considering HR-ARM, Gradient Boosting Classifier identified disorders of anal tone and contractility with an accuracy of 85.6% and an AUCROC of 0.910. Finally, disorders of the esophagogastric junction outflow were identified by Gradient Boosting Classifier with an accuracy of 94.2% and AUC-ROC of 0.921, while the xGBClassifier detected disorders of peristalsis with an accuracy of 80.9% and AUC-ROC of 0.871. Conclusion: In this study with an interoperable and multicentric design, the Gradient Boosting Classifier and xGBClassifier automatically identified motility disorders in ARM, HR-ARM and HREM. This study showcases the promise of AI-driven manometry, which could revolutionize functional gastrointestinal field trough increment in exam accessibility and accuracy.