Automated Multimodal Machine Learning for Esophageal Variceal Bleeding Prediction Based on Endoscopy and Structured...

Area of Science:

• Medical Imaging
• Machine Learning
• Gastroenterology

Background:

• Esophageal variceal (EV) bleeding is a life-threatening complication of cirrhosis.
• Early identification of high-risk patients is crucial for improved outcomes and resource allocation.

Purpose of the Study:

• To assess the feasibility of automated multimodal machine learning (MMML) for predicting 12-month EV bleeding risk.
• To integrate endoscopic images and clinical data for enhanced predictive accuracy.

Main Methods:

• Developed deep learning (DL) models using endoscopic images (ResNet, Xception, EfficientNet, ViT, ConvMixer).
• Trained and validated MMML models on an automated platform, combining DL outputs with clinical data (MELD, Child-Pugh, APRI, FIB-4).
• Externally tested MMML models and compared performance against existing clinical indices.

Main Results:

• An EfficientNet DL model achieved 0.868 accuracy in predicting bleeding events.
• A stacking MMML model demonstrated superior performance with 0.932 accuracy, 0.952 sensitivity, and 0.879 F1-score.
• The MMML approach outperformed traditional clinical indices in predicting 12-month EV bleeding.

Conclusions:

• Automated MMML is a feasible and effective tool for predicting 12-month EV bleeding risk.
• Integrating endoscopic imaging with clinical data significantly improves prediction accuracy.
• This novel approach offers a promising strategy for early risk stratification in cirrhotic patients.