Artificial intelligence for detection and analysis of motility patterns in high-resolution anorectal manometry: A multicentric multidevice transatlantic study

Abstract

Introduction Anorectal functional disorders are common in clinical practice, with highresolution anorectal manometry (HR-ARM) being the gold standard for their evaluation. However, HR-ARMislimitedbyits restrictedaccessibility and the complexity of data analysis. The London Protocol and Classification were developed to standardize the classification of anorectal motility patterns. This proof-of-concept study aims to develop and validate an artificial intelligence (AI) model for the identification and differentiation of disorders of tone and contractility in HR-ARM. Methods A dataset of 960 HR-ARM exams from two device manufacturers and three tertiary centers in Spain and the United States of America was utilized for the development and testing of multiple machine learning (ML) algorithms. The classification of anorectal motility disorders followed the London Classification. The exams were stratified into training and testing datasets using an 80%/20% patient split design based on label frequency. The models’ prediction was compared to a consensus-based classification by two HR-ARM experts (constituting the gold standard for HR-ARM analysis). The testing dataset was used to evaluate model performance through metrics as accuracy and the area under the receiver-operating characteristic curve (AUC-ROC). Model metrics represent the average of 10 independentevaluations.Results Afterfine-tuning, theGradientBoostingClassifier demonstrated the best overall performance, identifying disorders of anal tone and contractility with an accuracy of 85.6% and an AUC-ROC of 0.910. The LGBMClassifier detected anal normotension with hypocontractility with 90.5% accuracy and an AUC-ROC of 0.91. AdaBoostClassifier identified anal hypotension with hypocontractility with 94.9% accuracy and an AUC-ROC of0.944.Analhypotensionwithnormalcontractilitywasdetected by the LGBMClassifier with 94.1% accuracy and an AUC-ROC of 0.969. Finally, anal hypertension was identified by the GradientBoostClassifier with 96.9% accuracy and an AUC-ROC of 0.909. Discussion Various ML models accurately identified and differentiated disorders of anal tone and contractility in HR-ARM studies. The strong performance metrics observed across data from two device manufacturers address the challenge of interoperability, while the inclusion of data from two continents highlights the importance of mitigating demographic bias. AI models in HR-ARM have the potential to optimize exam accessibility and accuracy, significantly enhancing the management of patients with functional anorectal disorders.