An explainable machine learning model based on ultrasound and clinical features for impacted gallstone
TL;DR
The GBM model shows promise in identifying impacted gallstones preoperatively and a proof-of-concept web calculator was developed to explore individualized risk assessment, pending external validation.
Abstract
Impacted gallstones present a severe form of gallstones. Although positional adjustment during ultrasound examination has well-established clinical utility for evaluating stone impaction, there are still differences between the preoperative diagnosis results and the intraoperative findings. Machine-learning model may help accurate diagnosis prior to surgery. This retrospective cohort study included 1119 patients with gallbladder neck or duct stone. They underwent cholecystectomy to verify the stones were impacted or not. Fifteen candidate variables were evaluated, encompassing ultrasound features and clinical indicators. Six supervised algorithms were constructed with tenfold cross validation. Model performance was evaluated by discrimination, calibration, and decision curve analysis, and interpretability was assessed with Shapley additive explanations (SHAP). Nine variables were retained, including C-reactive protein, white blood cell count, Alkaline phosphatase, γ-glutamyltranspeptidase, neutrophil-to-lymphocyte ratio, gallbladder’s length and Length/Width, wall thickness, stone Length/Width. The GBM model exhibited the best performance with an AUC of 0.951 (95% CI: 0.930–0.973), and the lowest Brier score (0.1561) in the internal validation set. A proof-of-concept web calculator was developed to explore individualized risk assessment. The GBM model shows promise in identifying impacted gallstones preoperatively. Pending external validation, this tool may assist clinicians in making more precise decisions.