Predicting Timing of Surgical Intervention Using Recurrent Neural Network for Necrotizing Pancreatitis

Previous research focused on the qualitative discussion of the correlation between surgical time and mortality in acute pancreatitis (AP). Recommendations for surgical timing of necrotizing pancreatitis are delayed as far as possible, without recommendations for individuals. The aim of this article is to predict timing of surgical intervention in necrotizing pancreatitis with recurrent neural network (RNN). Time series data in AP were retrospectively extracted from a hospital in China (n = 15,813) to develop model, and the Cerner Health Facts database in the United States (n = 142,650) was used to externally validate. The developed model, time-aware Phased-Decay long short-term memory (LSTM), was used to predict timing of surgical intervention and critical clinical features in necrotizing pancreatitis based on laboratory tests, compared to other machine learning models. Area under the ROC curve (AUC) of RNN-based models was more than 0.70. The AUC of time-aware Phased-Decay LSTM (0.75) with more explanatory was similar to that (0.76) of Phased-Decay LSTM to predict surgical timing. The developed model visualized the specific surgical process and laboratory indicators changes for the patients with AP from the onset to discharge. Different clinical features had different contribution modes with time to the specific surgical event in AP. Heart function contributed the most to the prediction of surgical intervention at the onset and in the first week. Our developed model could monitor the specific surgical process from the onset of AP to discharge and extract the contribution modes of clinical features with time in necrotizing pancreatitis.