The Use of Bayesian Analysis to Predict Severe Acute Pancreatitis in a Nationwide Military Population

Abstract

Introduction: Predicting clinical outcomes is a valuable asset in improving patient care and optimizing medical resources. Acute pancreatitis (AP) is a relatively common gastrointestinal disease that can vary widely in its clinical severity. Bayesian network analysis is a method that may be useful to predict clinical outcomes of a variety of medical conditions. In this study, we used Bayesian network analysis to identify predictors of severe acute pancreatitis within our military health care system. Methods: Using data derived from the Military Healthcare System Data Repository (MDR), the total number of admissions, laboratory data, and clinical outcomes related to AP were examined from October 1, 2008 to September 30, 2012. Cases of AP were identified using ICD-9 codes associated with hospital admissions. Using a definition of a severe course as requiring ICU stay >48 hours, a Bayesian network analysis was applied to this cohort. Models were created using 80% of the data set then tested for accuracy with the remaining 20% of data. Multivariate logistic regression analysis was then performed to confirm these results. Results: A total of 3,134 cases of acute pancreatitis were identified out of a total study population of 2,973,523 for a cumulative incidence of 26 per 100,000. Six percent of these admissions (194 cases) were classified as severe AP. Bayesian network analysis identified 4 predictors of severe AP. The predictors included white blood cell count (WBC) >16,000 cells/mL, aspartate aminotransferase (AST) >250 IU/mL, serum creatinine (SCr) >2 mg/dL, serum calcium (Ca) <8 mg/dL (OR 4.15). These same 4 predictors were then confirmed by multivariate logistic regression analysis. When applied to the data set, the accuracy of the Bayesian model was 70.4%. Conclusion: Bayesian network analysis identified 4 clinical variables as predictors of severe AP and these were confirmed by a logistic regression analysis. Despite their agreement, the Bayesian network model had modest accuracy. Our results were limited by the inability to obtain certain important clinical data (i.e., radiographic findings, vital signs, mental status) from the electronic medical record, which have been established as clinically relevant predictors of AP. Despite these limitations, our results do indicate that Bayesian analysis is useful in predicting variables for a defined outcome. With a more complete data set, the accuracy of the Bayesian model will likely improve.