Validation of a machine learning model to non-invasively exclude spontaneous bacterial peritonitis

There remains a gap between guideline recommendations and real-world practice for the performance of timely paracentesis to diagnose spontaneous bacterial peritonitis (SBP). A machine learning (ML) model to non-invasively exclude SBP using twenty routinely collected clinical and laboratory values was previously developed and validated in a pre-COVID-19 era cohort.

Because of the considerable health care delivery changes post-COVID-19 pandemic, we sought to validate this ML model in a contemporary cohort of admitted VA patients.

We included patients in the Veterans Health Administration Corporate Data Warehouse (VHA-CDW) admitted between 2020 and 2023 with cirrhosis and ascites who underwent timely paracentesis identified by ICD-10 codes (validation cohort), then performed manual chart review on a subset of patients admitted at 2 tertiary-care VA hospitals to confirm the SBP (validation subgroup). We evaluated the performance metrics of the previously developed ML model with this cohort and subgroup.

The validation cohort included 4192 patients, of which 630 (15.0%) had SBP. At the <5%, <10%, and <15% thresholds for probability of SBP-negativity predicted by the ML model, negative predictive values (NPVs) were 93.6%, 92.0%, and 90.9%, respectively. In the validation subgroup, 7 (6.5%) of 107 patients had confirmed SBP. NPVs at the <5%, <10%, and <15% thresholds were 100%, 97.5%, and 95.4%, respectively.

In a national cohort and manual confirmation subgroup of Veterans with cirrhosis admitted after 2020, a ML model non-invasively excluded SBP with high NPV. This could help in risk-stratification, especially in limited-resource settings.