Margaret Brennan, n/a
Undergraduate Student
University of Rochester, United States
Disclosure information not submitted.
Robert Lindell, MD
Childrens Hospital of Philadelphia
Philadelphia, Pennsylvania
Disclosure information not submitted.
Akira Nishisaki, MD, MSCE
Associate Professor of Anesthesia and Critical Care Medicine
The Children's Hospital of Philadelphia, United States
Disclosure information not submitted.
Julie Fitzgerald, MD, PhD, FCCM
Children's Hospital of Philadelphia
Philadelphia, Pennsylvania
Disclosure information not submitted.
Adam Dziorny, MD, PhD,
Assistant Professor
Golisano Children's Hospital University of Rochester
Rochester, NY
Disclosure information not submitted.
Title: Application of Machine Learning in a Linked Multi-Center Dataset: Classifying MODS in Septic Shock
Introduction:
Pediatric sepsis is a leading cause of hospital mortality, and commonly causes multiple organ dysfunction syndrome (MODS). Manual identification of MODS is time-consuming and challenging to study in a single center due to heterogeneity. We describe a novel linked, multicenter database and demonstrate its use in identifying MODS using an optimized machine learning (ML) algorithm.
Methods:
We probabilistically linked subjects from the Virtual Pediatric System (VPS) and PEDsnet datasets using a validated method, and report descriptive statistics of the resulting linked dataset representing six academic institutions. We optimized, feature reduced, and re-trained a previously-described ML algorithm on a manually-reviewed MODS registry, and examined classification performance compared to the previous ML algorithm. We then applied this updated ML algorithm to severe sepsis patients from our linked dataset, and identified cumulative incidence of MODS on days 1-7 as well as all-cause mortality and PICU length of stay (LOS).
Results:
Among 82,457 patients identified in VPS as admitted between 2012–2017 to six academic centers, 78,856 (96%) were probabilistically linked to PEDSnet patients. From this dataset, 5,426 (6.9%) had septic shock by VPS diagnoses. A previously-described ML model was reduced to 6 features and re-trained using 10-fold cross-validation across 652 patient-days. Classification of MODS, compared to previous model, yielded an improved sensitivity of 0.92 (prior: 0.73), specificity of 0.81 (prior: 0.86), and improved PPV of 0.90 (prior: 0.82). Features included invasive respiratory procedures, # of cardiac stimulant medications, mean pH, max BUN, mean ALT, and min platelets. Application to severe sepsis patients resulted in cumulative MODS incidence of 49%. Patients with MODS had higher PRISM III scores, higher mortality (11.3% vs 1.1%, p< 0.001), and longer median PICU LOS (7.2 vs 1.9 days, p< 0.001).
Conclusions:
An optimally tuned ML model can classify MODS based on reduced features consistent with individual organ system dysfunction. Using our ML model, we have constructed a multicenter cohort of accurately phenotyped patients with sepsis and MODS, suitable for comparative effectiveness research focused on the association between adjunctive therapies and outcomes.