Multimodal predictive model for strangulation risk in adhesive small bowel obstruction using deep learning and...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Adhesive small bowel obstruction (ASBO) poses a significant risk of strangulation.
• Accurate prediction of strangulation is crucial for timely intervention and improved patient outcomes.
• Current diagnostic methods have limitations in predicting strangulation risk.

Purpose of the Study:

• To develop and validate a multimodal predictive model for strangulation risk in ASBO.
• To integrate deep learning-based computed tomography (CT) imaging features with clinical electronic health records (EHR).
• To enhance diagnostic accuracy and support clinical decision-making in ASBO management.

Main Methods:

• Retrospective, observational, multicenter study using data from 225 patients for development and 123 for validation.
• Utilized a 3D convolutional neural network (ResNet50) for CT image analysis and strangulation risk classification.
• Integrated deep learning predictions with top EHR features using the XGBoost algorithm for a multimodal model.

Main Results:

• The multimodal model achieved superior performance in predicting strangulation within 7 days (AUC 0.915 training, 0.912 testing).
• Demonstrated excellent calibration, significant clinical utility via decision curve analysis, and improved prediction with deep learning.
• Outperformed single-modality models in predicting strangulation risk.

Conclusions:

• Multimodal artificial intelligence integrating imaging and clinical data shows significant potential for improving ASBO strangulation risk prediction.
• The developed model can enhance diagnostic accuracy and aid clinical decision-making in ASBO.
• This approach offers a promising tool for better patient management and outcomes in ASBO.