STARK STUDY: MACHINE LEARNING APPROACH TO PREDICT POST-ERCP PANCREATITIS IN AN INTERNATIONAL MULTICENTER PROSPECTIVE COHORT STUDY

Aims Post-endoscopic retrograde cholangiopancreatography(ERCP) acute pancreatitis(PEP) is ERCP most frequent complication. Predicting PEP onset risk can be determinant in reducing its incidence. However, studies conducted so far identified single risk factors that have never been used together to predict the risk effectively. The aim of our study was to build a mathematical model to predict PEP probability through machine learning techniques.

Methods “STARK project” is an international, multicenter,prospective cohort study developed within Pancreas 2000 Educational Program, carried out in 7 tertiary centers enrolling patients undergoing ERCP. Patients enrolled were followed-up to detect PEP.

The data was randomly split in training set(80%) and test set(20%). Two models were used to predict PEP probability: gradient boosting(GB) and logistic regression(LR). On both models the same data preparation and the same following procedure was applied:on the training set, a 10-split random cross-validation(CV) was applied to optimize parameters in order to obtain the best mean Area Under the Curve(AUC). Afterwards, the model was re-trained on the whole training set with the best parameters and then applied on the test set.

Results 1,150 patients were included. 70(6.1%) patients developed PEP. Model most relevant variables for the prediction of PEP were:total bilirubin level, body mass index,age, procedure time, units of alcohol/day, previous sphincterotomy.

GB model retrieved a ROC AUC in CV of 0.699±0.076 with 95% CI 0.64-0.76; ROC AUC in test was 0.671. LR model retrieved a ROC AUC in CV of 0.584±0.068 with 95% CI 0.58-0.63; ROC AUC in test was 0.555.

The statistical comparison between the two models in CV retrieved a p value of 0.01.

Conclusions This is the first study applying machine learning techniques for the prediction of PEP, with the GB model showing a significantly better performance than the LR model. The most relevant variables we observed were mostly pre-procedural variables except for the procedure time.