A machine learning-based predictive model for biliary stricture attributable to malignant tumors: a dual-center retrospective study.

BACKGROUND

Biliary stricture caused by malignant tumors is known as Malignant Biliary Stricture (MBS). MBS is challenging to differentiate clinically, and accurate diagnosis is crucial for patient prognosis and treatment. This study aims to identify the risk factors for malignancy in all patients diagnosed with biliary stricture by Endoscopic Retrograde Cholangiopancreatography (ERCP), and to develop an effective clinical predictive model to enhance diagnostic outcomes.

METHODOLOGY

Through a retrospective study, data from 398 patients diagnosed with biliary stricture using ERCP between January 2019 and January 2023 at two institutions: the First People's Hospital affiliated with Jiangsu University and the Second People's Hospital affiliated with Soochow University. The study began with a preliminary screening of risk factors using univariate regression. Lasso regression was then applied for feature selection. The dataset was divided into a training set and a validation set in an 8:2 ratio. We analyzed the selected features using seven machine learning algorithms. The best model was selected based on the Area Under the Receiver Operating Characteristic (ROC) Curve (AUROC) and other evaluation indicators. We further evaluated the model's accuracy using calibration curves and confusion matrices. Additionally, we used the SHAP method for interpretability and visualization of the model's predictions.

RESULTS

RF model is the best model, achieved an AUROC of 0.988. Shap result indicate that age, stricture location, stricture length, carbohydrate antigen 199 (CA199), total bilirubin (TBil), alkaline phosphatase (ALP), (Direct Bilirubin) DBil/TBil, and CA199/C-Reactive Protein (CRP) were risk factors for MBS, and the CRP is a protective factor.

CONCLUSION

The model's effectiveness and stability were confirmed, accurately identifying high-risk patients to guide clinical decisions and improve patient prognosis.