LAUSR

Introduction/Hypothesis: Pediatric severe sepsis is a common cause of multiple organ dysfunction syndrome (MODS). Current methods to identify MODS are time-consuming and require extensive chart review. We hypothesize that machine learning (ML) algorithms trained on a manuallyreviewed MODS registry can accurately classify MODS through day 7 of illness. Methods: IRB-approved single-center retrospective study of PICU patients (1/2012–12/2017). Probabilistic linkage model joined patients in VPS (Virtual Pediatric Systems) and PEDSnet datasets and selected for diagnoses of severe sepsis or septic shock (SS). Features were chosen a priori based on MODS definitions and features available in linked dataset, including vital signs, laboratory values, medications, procedures, and diagnoses. ML pipelines generated using TPOT (Tree-based Pipeline Optimizer Tool) trained models against gold-standard sepsis registry with manually abstracted MODS on days 1 through 7 on data split 75%/25%. Confusion matrices and standard test characteristics were calculated for MODS models, as well as for individual organ system dysfunction and day of illness. Results: Linked dataset included 1,250 PICU admissions with SS. After feature extraction, preprocessing, and accounting for mortality and discharges, 684 patient-day rows were available for analysis with 187 total features. Using 8-fold cross-validation, classification of MODS (yes/no) yielded an overall sensitivity=73%, specificity=85%, PPV=81%. Important features include # of cardiac stimulant medications (ATC class C01C), # blood gases, VPS respiratory disease diagnoses, liver function test values, pH, sodium and systolic BP. Feature selection and recursive feature elimination did not substantially alter results, nor did subset analyses by day of illness. In the predicted dataset, mortality was 13.3% in patients with MODS predicted on any day, compared to 3.8% in non-MODS patients. Conclusions: Optimized machine learning algorithms can classify MODS beyond day 1 of illness in a linked dataset which leverages the accurate diagnosis codes of VPS and the granular measurements of PEDSnet. Application of this technique to a multi-center linked dataset will create a large epidemiologic database to study sepsis-induced MODS and clinical outcomes in an accurately-defined patient population.