749 - The Implementation of Machine Learning to Predict Pediatric Inpatient Deterioration

Monday, May 1, 2023

9:30 AM – 11:30 AM ET

Poster Number: 749
Publication Number: 749.414

Michael W. Stevens, University of Michigan Medical School, Ann Arbor, MI, United States; Brandon Cummings, University of Michigan Medical School, Ann Arbor, MI, United States; Sardar Ansari, University of Michigan Medical School, Ann Arbor, MI, United States; Maureen Brudzinski, University of Michigan Medical School, Ann Arbor, MI, United States; Claire Kalpakjian, University of Michigan Medical School, Ann Arbor, MI, United States; Rodney C. Daniels, University of Michigan, Ann Arbor, MI, United States

Presenting Author(s)

Michael W. Stevens, MD (he/him/his)

Pediatric Critical Care Fellow
University of Michigan Medical School
Ann Arbor, Michigan, United States

Background: A challenge of inpatient medicine is predicting which patients will decompensate. P</b>redicting Intensive Care Transfers and UnfoReseen Events - pediatrics (PICTURE-Peds) is a machine learning analytic developed to help identify pediatric patients at risk for clinical deterioration. Integration of this technology into healthcare system workflows requires a multidisciplinary approach to ensure uptake.

Objective: In this project, we used a qualitative methodology to examine potential barriers and facilitators of uptake of PICTURE among inpatient clinicians. These data are essential to guide the introduction of PICTURE in the clinical workflow to support clinician decision making. These insights will serve as the backbone of future sessions that will develop the interface of PICTURE-Peds.

Design/Methods: Our team included computer/data scientists, human centered designers, a biopsychosocial scientist, and a pediatric critical care faculty and fellow. PICTURE was validated using retrospective and prospective data collection. Then, to ensure PICTURE operated with optimal efficiency and user impact, we conducted a series of semi-structured interviews with stakeholders. We used an inductive and deductive approach to explore aspects of a computerized early warning system that would be of value to clinicians. The interviewer was a third-year pediatric critical care fellow with board certification in internal medicine and general pediatrics. We used thematic analysis to analyze interview data; this approach offers a structed framework of pattern identification and theme development using inductive and deductive reasoning.

Results:
Eight clinicians (3 chief residents, 2 nursing supervisors, 2 nurses and 1 hospitalist) participated in interviews. Five themes emerged including data presentation, communication, salient information, system frequency, and transparency. These themes informed specific design concepts to facilitate adoption: PICTURE should alert all team members to improve communication; alerts should provide novel data; it should minimize gaps in patient assessment; patient data should be visualized across time; and the data behind its validation should be transparent.

Conclusion(s):
Results of interviews highlighted the importance of PICTURE providing new and actionable information in a cohesive and efficient way to the medical team with the data PICUTRE used visible. With this foundation, the next step in development is to conduct a series of design sprints to generate the user interface and other features of PICTURE to deploy in the electronic medical record.