Development and Validation of an Explainable Machine Learning Model for Gangrenous Cholecystitis Prediction: A...

Area of Science:

• Medical Informatics
• Surgical Oncology
• Machine Learning in Healthcare

Background:

• Gangrenous cholecystitis (GC) requires prompt diagnosis and surgical intervention.
• Accurate preoperative risk stratification for GC is crucial for patient management.
• Existing diagnostic methods may have limitations in early GC detection.

Purpose of the Study:

• To develop and externally validate an interpretable machine learning (ML) model for predicting gangrenous cholecystitis (GC) preoperatively.
• To utilize multicenter clinical data for robust model development and validation.
• To enhance the transparency of ML model decisions in predicting GC.

Main Methods:

• Retrospective analysis of 744 patients with cholecystitis, split into training and testing cohorts, with external validation on 300 patients.
• Feature selection using LASSO regression and Boruta algorithm from 20 preoperative variables.
• Construction and evaluation of six ML models using AUC, calibration, and decision curve analysis, with SHAP for interpretability.

Main Results:

• The Random Forest (RF) model achieved superior performance: AUC of 0.893 (training), 0.875 (testing), and 0.818 (external validation).
• Excellent calibration and clinical benefit demonstrated by decision curve analysis.
• Key predictors identified by SHAP analysis include gallbladder wall thickening, C-reactive protein, pericholecystic fluid, white blood cell count, and impacted stone.

Conclusions:

• An interpretable ML model demonstrates good discrimination and generalizability for preoperative GC risk stratification across multicenter cohorts.
• The model integrates clinical, laboratory, and imaging features, offering explainability to aid perioperative decision-making.
• Further prospective multicenter evaluations are recommended before routine clinical adoption.