O1 Using Machine Learning to predict the Severity of Acute Pancreatitis: A Systematic Review

Abstract Background Acute pancreatitis (AP) is an inflammatory condition of the pancreas where, in its severe form, has a mortality of up to 30%. The severity of AP is traditionally predicted with using static scoring systems such as Ranson and APACHE II based on clinical observations and/or blood test results. The Revised Atlanta Classification of AP, however, include the assessment of local complications such as peripancreatic fluid collections and necrosis which can be more subjective and relies on human clinical or radiological judgement. The aim of this study is to review the literature and evaluate the potential role and effectiveness of machine learning (ML) in the stratification of AP severity. Methods A systematic review was conducted using PubMed with search terms including acute pancreatitis, severity and machine learning. The inclusion criteria selected studies describing the development and validation of ML models predicting AP severity and to those publishing sensitivity and specificity data. Results There were 97 articles identified which after applying the inclusion and exclusion criteria, 11 studies were included in the review. ML training datasets ranged from 265 to over 350,000 patients, with inputs varying from patient demographics, clinical observations and blood results as well as radiomic data. The highest sensitivity (96%) and specificity (98%) predictors for severe AP outperformed Ranson, APACHE II and BISAP scoring systems. Conclusion ML models display superior accuracy when compared to traditional scoring systems in predicting AP severity. Radiomic data from Computed Tomography images can increase the sensitivity and specific of predictive models.