16 Model for Predicting Severity in Acute Pancreatitis

Abstract

INTRODUCTION: Acute pancreatitis (AP) is severe in about 20% of cases with an associated mortality as high as 40%. Accurate prediction of AP severity is necessary to both triage patients and to identify those that might require specific early medical management (EMM). Existing AP severity scoring methods are based on retrospective data and derived from cohorts managed without standardized EMM. The aim of this study was to identify a prognostic model to predict AP severity in a prospective cohort using an evidence-based EMM protocol. METHODS: All patients diagnosed with AP from 1/2015–9/2018 were managed per protocol that included emergency department and admit order sets to standardize EMM, e.g., early aggressive fluid resuscitation and enteral feeding. Exclusion criteria included age less than 18 yrs, transfers from outside hospitals, post-ERCP AP, and trauma related AP. Final severity of AP was categorized as either mild or severe (severe and moderately severe) as per Revised Atlanta Classification. Univariate analysis was performed to identify variables of interest. Classification and regression tree (CART®) analysis (a non-parametric method based on the binary recursive partitioning of data) was used to identify the baseline variables most important in predicting severity, determine cutoffs, and to create a decision tree model. Class probability function was used as the splitting rule for tree building with the limit set to not split the node if sample <30 and that the terminal node not have sample <10. We used 20-fold cross validation method for validating the tree model, the overall performance of which was validated in the test sample using area under the receiver operating characteristic (ROC) curve. RESULTS: 417 total patients were included in the cohort; 356 and 61 patients met criteria for mild and severe pancreatitis, respectively. CART analysis identified five baseline variables in order of greatest to least importance: Creatinine (Cr), procalcitonin, SIRS, age, and HCT (Figure 1). The analysis identified a prognostic algorithm using cutoffs to predict AP severity (Figure 2). ROC curves were generated for the learned (83.7%) and test (73.6%) data sets, respectively (Figure 3). CONCLUSION: We identified baseline variables and an algorithm to predict AP severity from a prospective cohort managed using a standardized, evidence based approach. Our results should be validated prospectively to generate a simple scoring model that can be used to appropriately care for patients with AP.