Construction of a Multi-View Deep Learning Model for the Severity Classification of Acute Pancreatitis

Background: Acute pancreatitis (AP) is a prevalent pathological condition of abdomen characterized by sudden onset, high incidence and complex progression. Timely assessment of AP severity is crucial for informing intervention decisions so as to delay deterioration and reduce mortality rates. Existing AP-related scoring systems can only assess current condition of patients and utilize only a single type of clinical data, which is of great limitation. Therefore, it is imperative to establish more accurate and data-compatible methods for predicting the severity of AP. The artificial intelligence (AI) algorithm based on artificial neural network (ANN) allow for the adaptive feature extraction for objective task through its internal complex network, instead of the hand-crafted methods commonly used in traditional machine learning (ML) algorithms. In this study, we delve into the final severity classification prediction of newly admitted AP patients, using deep learning (DL) algorithm to develop multi-view models, incorporated with patients' demographic information, vital signs, AP-related laboratory indexes and admission computed tomography (CT) images.

Methods: The pancreatitis database in the platform of Clinical Data Research Center of Acute Abdominal Surgery at the First Affiliated Hospital of Dalian Medical University was used to gather AP cases. Deep neural network (DNN) and convolutional neural network (CNN) were utilized to construct models. The DNN prediction models with clinical data as input, the CNN prediction models with admission CT as input, and the multi-view models combining the two inputs were respectively established to predict the severity of AP.

Results: DL models for AP severity classification based on clinical indexes, imaging data and merged data were constructed. The multi-view model based on merged data offered more accurate prediction of the final severity classification of AP, with an overall accuracy rate of 80.26% (95% confidence interval (CI): 79.58%-80.94%). The constituent accuracy rates for mild acute pancreatitis, moderately severe acute pancreatitis, and severe acute pancreatitis were 91.69% (95% CI: 87.80%-95.57%), 64.90% (95% CI: 58.85%-70.95%), and 75.56% (95% CI: 68.58%-82.53%), respectively.

Conclusion: The multi-view models using clinical indexes and imaging data as input outperform single-view models in AP severity prediction.