Machine Learning Model In Acute Pancreatitis: Overcoming the Limitations with Prognostication Models

Abstract

Introduction: Prognostic scoring systems (PSS) play an important role in managing patients with acute pancreatitis (AP). Existing scores have been used and studied extensively, and guides clinical management based on predicted severity. However, traditional PSSs may have reached their limits. This study aims to investigate a machine learning (ML) model for prognosticating AP. Methods: All patients diagnosed with acute pancreatitis at the National University Hospital (NUH), Singapore, were recruited from 2012 to 2017. Severity of AP was determined using Revised Atlanta Criteria. Patients with mild AP were placed in category 1, while those with moderate or severe AP were in category 2. Patient data available within 2 hours of patient presentation was used to develop the ML model using supervised learning. Missing data was accounted for using K-nearest neighbour. Statistical analysis was performed using SPSS 26.0. Results: 599 patients were recruited. Existing PSSs such as HAPS, Glasgow, Ranson’s, BISAP and SIRS were evaluated (Table 1). 72 variables were included in a Random Forest ML model giving a sensitivity of 78%, specificity of 100%, AUC of 0.89, accuracy of 96%, and Matthews correlation coefficient (MCC) of 0.86 for prediction of AP severity. Variables with the heaviest weightage were serum creatinine, urea, and PaO2/FiO2 ratio. Conclusion: PSSs are important for triage, resource allocation, early discharge and early cholecystectomy. ML based prognostic algorithms may be the way forward in developing a PSS with higher degrees of accuracy and certainty. ML algorithms also enable the models to continually learn and improve on itself.Tabled 1Scoring SystemSensitivitySpecificityPPVNPVHAPS41.3%74.6%85.1%26.5%Glasgow77.0%55.1%85.1%41.8%Ranson76.9%45.3%82.4%36.9%BISAP96.7%17.1%80.3%59.5%SIRS72.5%35.8%79.5%27.6% Open table in a new tab