ارزو رحيم زاده كيوي

Endoscopic Retrograde Cholangiopancreatography (ERCP) is the primary therapeutic tool for biliary an‎d pancreatic duct disorders. Despite its high efficacy, the most common complication of this complex procedure is Post-ERCP Pancreatitis (PEP), which has an incidence rate of 3.5-9.7% an‎d a mortality rate of 0.7%, posing a threat to patient health, increasing hospital stays, an‎d raising treatment costs. Therefore, the timely identification of high-risk patients can play a significant role in preventing this complication. The multitude of risk factors an‎d their non-linear interactions limit the effectiveness of traditional PEP prediction methods. To address this, the present study developed a PEP prediction model utilizing advanced machine learning algorithms an‎d metaheuristic feature selec‎tion methods. Real-world data from 1026 patients in Golstan province, collected from three medical centers between 2017 an‎d 2024, were used. Six machine learning algorithms—Logistic Regression, K-Nearest Neighbors, Support Vector Machine, Ran‎dom Forest, Light Gradient Boosting Machine, an‎d Extreme Gradient Boosting—were eva‎luated within a nested cross-validation framework an‎d with an independent external test set. The use of feature selec‎tion methods, namely Particle Swarm Optimization an‎d Genetic Algorithm with improved initial solutions, led to a significant enhancement in the balance between sensitivity an‎d specificity an‎d the F1-score of the models. Among them, Particle Swarm Optimization combined with Light Gradient Boosting demonstrated the best performance on the external test set, achieving an F1-score of 0.81 an‎d an AUC of 0.96. Additionally, Ran‎dom Forest, utilizing only 13 features selec‎ted by the Genetic Algorithm from the original 59, showed good performance with an F1-score of 0.76. Finally, a stacking ensemble model based on the top three base models, optimized with a calibrated threshold, achieved a superior performance with an F1-score of 0.83 an‎d an AUC of 0.95. Difficult cannulation, pancreatic duct cannulation, pancreatic cancer, younger age, female sex, an‎d specific papillary morphology were identified as the most influential predictors. This study is the first in Iran to propose a PEP prediction model by integrating multi-center data, metaheuristic algorithms, an‎d a rigorous eva‎luation framework. The results can be utilized as a clinical decision support system for the early identification of high-risk patients.

پانكراتيت پس از ERCP , يادگيري ماشين , انتخاب ويژگي فراابتكاري , مدل تركيبي , عامل خطر

Post-ERCP Pancreatitis , Machine learning , Metaheuristic feature selec‎tion , Ensemble model , Risk factor

Arezoo Rahimzadeh Kivi

Farnaz Barzinpour