Scientific Reports (May 2022)

Early prediction of acute necrotizing pancreatitis by artificial intelligence: a prospective cohort-analysis of 2387 cases

  • Szabolcs Kiss,
  • József Pintér,
  • Roland Molontay,
  • Marcell Nagy,
  • Nelli Farkas,
  • Zoltán Sipos,
  • Péter Fehérvári,
  • László Pecze,
  • Mária Földi,
  • Áron Vincze,
  • Tamás Takács,
  • László Czakó,
  • Ferenc Izbéki,
  • Adrienn Halász,
  • Eszter Boros,
  • József Hamvas,
  • Márta Varga,
  • Artautas Mickevicius,
  • Nándor Faluhelyi,
  • Orsolya Farkas,
  • Szilárd Váncsa,
  • Rita Nagy,
  • Stefania Bunduc,
  • Péter Jenő Hegyi,
  • Katalin Márta,
  • Katalin Borka,
  • Attila Doros,
  • Nóra Hosszúfalusi,
  • László Zubek,
  • Bálint Erőss,
  • Zsolt Molnár,
  • Andrea Párniczky,
  • Péter Hegyi,
  • Andrea Szentesi,
  • Hungarian Pancreatic Study Group

DOI
https://doi.org/10.1038/s41598-022-11517-w
Journal volume & issue
Vol. 12, no. 1
pp. 1 – 11

Abstract

Read online

Abstract Pancreatic necrosis is a consistent prognostic factor in acute pancreatitis (AP). However, the clinical scores currently in use are either too complicated or require data that are unavailable on admission or lack sufficient predictive value. We therefore aimed to develop a tool to aid in necrosis prediction. The XGBoost machine learning algorithm processed data from 2387 patients with AP. The confidence of the model was estimated by a bootstrapping method and interpreted via the 10th and the 90th percentiles of the prediction scores. Shapley Additive exPlanations (SHAP) values were calculated to quantify the contribution of each variable provided. Finally, the model was implemented as an online application using the Streamlit Python-based framework. The XGBoost classifier provided an AUC value of 0.757. Glucose, C-reactive protein, alkaline phosphatase, gender and total white blood cell count have the most impact on prediction based on the SHAP values. The relationship between the size of the training dataset and model performance shows that prediction performance can be improved. This study combines necrosis prediction and artificial intelligence. The predictive potential of this model is comparable to the current clinical scoring systems and has several advantages over them.