A MACHINE LEARNING ALGORITHM GENERATED USING ROUTINE PARAMETERS COULD AVOID UNNECESSARY TAPS TO EXCLUDE SBP IN A NATIONAL VA COHORT WITH INTERNAL AND EXTERNAL VALIDATION

Background and Aims: Despite the poor prognosis of missed/delayed diagnosis of spontaneous bacterial peritonitis (SBP), there are logistical challenges in providing timely tap. These are related to expertise, training, and time needed in busy emergency settings. Therefore, measures to exclude SBP that obviate taps are needed. Aim: Use machine learning (ML) to create and validate models to exclude SBP without a tap using easily observed and collected parameters.
Method: Training cohort: Using VA Corporate Data Warehouse (CDW), records of patients with cirrhosis and ascites who underwent first paracentesis within a day of admission between 2009-2019 were included. SBP was defined as PMN>250/ml, positive culture, or validated SBP billing code. Demographics, admission laboratories, medications, vital signs & comorbidities were recorded. XGBoost analysis was performed to determine the probability of excluding SBP based on variables that did not include ascites fluid analysis. The generated XGBoost model was tested in 2 validation cohorts: (a) in CDW: 2nd para during 2009-2019 and (2) Prospective NACSELD cohort from VCU/Richmond VA admitted non-electively (Fig 1). Negative predictive value (NPV) at 5,10 & 15% probability were tested.
Results: Training cohort: 9,643 patients (63.14±8.69 years age,95.5% men, 50.7% alcohol-related) got their first early tap, with 1,448 (15.0%) having SBP. Model was trained on 75% of the cohort and 25% was reserved for testing. We initially collected 147 predictor variables, then reduced to top-20 variables selected by XGBoost (Fig 2A) with test set NPV for SBP 96.5% (Fig 2B). These included demographics (BMI), vitals (temp, BP) and labs (liver function tests, basic panel, CBC) at the time of admission.
Validation cohort #1: 2844 pts (63.14±8.37 years age, 97.4% men, 54.0% alcohol-related) had repeat paracentesis of which 10% had SBP. Using the 20-variable ML model, NPV to exclude SBP at all levels was high (Fig 2C). Validation cohort #2: 333 pts from NACSELD (57.08±7.74years age, 65.1%men, 50.1% alcohol-related) were admitted non-electively of whom 9% had SBP on tap. Fig 2C shows high NPV using the ML model.
Conclusion: A machine learning model to exclude SBP without needing a tap using 20 variables available at time of admission (demographics, labs, and vitals) showed excellent NPV in a national VA Veterans cohort. The high NPV for SBP were validated in internal and external cohorts and could be utilized in emergency settings to reliably identify which patient with cirrhosis and ascites who can avoid a tap to exclude SBP.