Prediction of post-operative acute pancreatitis in children with pancreaticobiliary maljunction using machine learning model

Purpose: This study aimed to develop a prediction model to identify risk factors for post-operative acute pancreatitis (POAP) in children with pancreaticobiliary maljunction (PBM) by pre-operative analysis of patient variables.

Methods: Logistic regression (LR), support vector machine (SVM), and extreme gradient boosting (XGBoost) models were established using the prospectively collected databases of patients with PBM undergoing surgery which was reviewed in the period comprised between August 2015 and August 2022, at the Children's Hospital of Soochow University. Primarily, the area beneath the receiver-operating curves (AUC), accuracy, sensitivity, and specificity were used to evaluate the model performance. The model was finally validated using the nomogram and clinical impact curve.

Results: In total, 111 children with PBM met the inclusion criteria, and 21 children suffered POAP. In the validation dataset, LR models showed the highest performance. The risk nomogram and clinical effect curve demonstrated that the LR model was highly predictive.

Conclusion: The prediction model based on the LR with a nomogram could be used to predict the risk of POAP in patients with PBM. Protein plugs, age, white blood cell count, and common bile duct diameter were the most relevant contributing factors to the models.