Predicting Neoplastic Polyp in Patients With Gallbladder Polyps Using Interpretable Machine Learning Models:...

Area of Science:

• Gastroenterology and Hepatology
• Medical Informatics
• Oncology

Background:

• Gallbladder polyps (GBPs) are common, with a subset posing a risk of malignant transformation.
• Accurate differentiation between benign and neoplastic GBPs is crucial for appropriate management.
• Neoplastic GBPs require timely intervention to prevent progression to gallbladder cancer.

Purpose of the Study:

• To develop and validate interpretable machine learning (ML) models for predicting neoplastic GBPs.
• To identify key features predictive of neoplastic transformation in GBPs.
• To enhance model transparency and clinical utility using Shapley additive explanations (SHAP).

Main Methods:

• Retrospective analysis of 924 patients with GBPs who underwent cholecystectomy.
• Utilized patient characteristics, lab results, ultrasound, and pathology data.
• Developed and compared nine ML algorithms, evaluating performance with AUC and SHAP for interpretability.

Main Results:

• K-nearest neighbors, C5.0 decision tree, and gradient boosting machine models demonstrated superior predictive performance.
• The SHAP method identified key predictors, with polyp size being the most significant.
• Lesions ≥18mm were highlighted as requiring heightened clinical surveillance and prompt intervention.

Conclusions:

• Interpretable ML models offer accurate prediction of neoplastic GBPs.
• These models aid in treatment planning and resource allocation for GBP patients.
• Model transparency builds physician trust, facilitating confident application in patient care.