Construction and validation of a prediction model for textbook outcome in distal cholangiocarcinoma undergoing pancreaticoduodenectomy based on six machine learning algorithms.

BACKGROUND: Textbook outcome (TO) serves as a validated composite metric for short-term surgical quality in pancreaticoduodenectomy (PD) patients. This study developed and validated a machine learning-based model to predict the probability of TO attainment in patients with distal cholangiocarcinoma (dCCA) undergoing PD.

METHODS: Patients with dCCA who underwent PD at our institution between 2012 and 2020 were retrospectively analyzed. Candidate variables were first screened using least absolute shrinkage and selection operator (LASSO) regression, and the selected features were subsequently entered into multivariable logistic regression to identify independent predictors of TO. Six machine learning models were then constructed and evaluated, with their performance compared using the area under the receiver operating characteristic curve (AUC-ROC). SHapley Additive exPlanations (SHAP) analysis was used to elucidate the optimal model's decision-making process and quantify the contribution of each predictor.

RESULTS: Among the 242 patients included in the study, 177 (73.1%) achieved TO. LASSO regression followed by multivariable logistic analysis identified five independent predictors of TO: ASA classification, lymph node metastasis, perineural invasion, BMI, and operation time. Six machine learning algorithms were constructed, among which the Gradient Boosting Machine (GBM) exhibited the best discriminative performance (AUC training set: 0.951, 95% CI 0.915-0.988; test set: 0.890, 95% CI 0.800-0.980). Calibration plots and decision curve analysis further demonstrated the strong accuracy and favorable clinical utility of the GBM-based model.

CONCLUSION: The GBM-based prediction model offers accurate preoperative probability estimates of TO attainment and serves as a valuable tool to support individualized surgical planning.