Proposal Summary There are no validated systems for identifying children without serious bacterial infection (SBI) upon admission to a pediatric ICU (PICU). Given the high prevalence of SBI among critically ill children (up to 46%) and risks associated with delayed antibiotic administration, nearly 50% of children without SBI receive antibiotics while microbiologic studies are pending. However, antibiotics can have adverse effects including acute kidney injury, clostridium difficile colitis, and development of antibiotic resistance. The long-term goal of this research is to validate and disseminate machine learning (ML)-based clinical decision support (CDS) tools able to improve PICU antibiotic decision-making thereby reducing antibiotic associated harm among critically ill children. In prior work, Dr. Martin developed ML-based predictive models, which use electronic health record (EHR) inputs (vital sign, laboratory, and other clinical data), to accurately identify children without SBI upon PICU admission in a single center retrospective cohort. The central hypothesis is that these models will demonstrate similar robust performance during prospective and multicenter evaluations, and that an antibiotic decisional needs analysis of PICU clinicians will inform the optimal design of model-based CDS tools. The central hypothesis will be tested via three aims: 1) prospectively evaluate two SBI predictive models within a single center EHR and determine the potential effect on antibiotic-days per child; 2) evaluate ML model generalizability by testing them in a multicenter EHR cohort; and 3) perform a multicenter, multidisciplinary antibiotic decisional needs analysis of PICU clinicians to facilitate user-centered design of equitable model-based CDS tools. In Aim 1, two SBI predictive models will be prospectively evaluated in silent fashion (predictions not revealed to clinicians) at a single center over two years. Model predictions will be compared to patient SBI outcomes to determine their negative predictive value and potential to reduce unnecessary antibiotics. In Aim 2, the same models will be applied to a retrospective dataset of six US children's hospital PICUs (~178,000 encounters over 8+ years) to assess generalizability by determining each model's negative predictive value and potential to reduce unnecessary antibiotics. In Aim 3, a rigorous qualitative content analysis of PICU clinician interviews from five institutions will identify the values driving antibiotic decision-making and enable user-centered design of model- based CDS tools. The research is innovative because it involves development of the first clinically validated system for excluding SBI at PICU admission and uses a ML approach to do so. The research is significant as it accelerates development of generalizable antibiotic decision-making tools to assist PICU clinicians in safely minimizing unnecessary antibiotics and associated harm. The educational component of this applicatio...