Accurate and Reliable Diagnostics for Injured Children: Machine Learning for Ultrasound

PROJECT SUMMARY/ABSTRACT Dr. Aaron Kornblith, a general and pediatric emergency physician at the University of California, San Francisco (UCSF) is establishing himself as a future investigator in patient-oriented clinical research of novel diagnostics in injured children. This award will enable him to accomplish the following goals: (1) become an expert at patient-

oriented clinical research in pediatric abdominal trauma; (2) develop novel machine learning models for a bedside ultrasound application; (3) implement advanced computational methods to develop, validate, and test clinical decision rules incorporating bedside ultrasound; and (4) develop an independent clinical research career.

To achieve these goals, Dr. Kornblith has assembled an expert mentoring team: primary mentor Dr. Jeffrey Fineman, Chief of Pediatric Critical Care at UCSF (conducts clinical investigations in children with critical illness and is an expert in career development of early-stage investigators), co-mentors Dr. Atul Butte, (an expert in

healthcare and data science), Drs. James Holmes and Nathan Kuppermann (experts in the diagnostic evaluation of pediatric trauma and clinical decision rules), scientific advisor Dr. John Mongan, (expert in developing, validating, and implementing machine learning for imaging tasks), and statistical advisor Dr. Bin Yu (an expert

in statistical theory including accurate, reliable, and interpretable computational methods, and implicit bias). Hemorrhage from blunt intraabdominal injury is a leading cause of death in children. Identifying abdominal hemorrhage early is essential to minimizing morbidity and mortality from delayed or missed diagnoses. The

reference standard test, abdominal computed tomography (CT), has drawbacks including risk of radiation- induced malignancy. For 25-years, CT use in children has increased dramatically without proportional improvements in outcomes. Focused Assessment with Sonography for Trauma (FAST) is a bedside ultrasound

method to evaluate children for abdominal hemorrhage. FAST may help clinicians balance the risk of missed intraabdominal injury with unnecessary exposure to ionizing radiation from CT. Dr. Kornblith’s research will focus on improving pediatric FAST’s accuracy and reliability using machine learning models (Aim 1) and

developing/validating novel clinical decision rules incorporating FAST to identify children at very low risk for injury who can forgo CT (Aim 2). Dr. Kornblith will use an existing dataset and computing infrastructure to develop and validate a machine learning model using >2.1 million frames from 1,264 pediatric FAST studies to detect

hemorrhage as accurately as an expert (Aim 1), and two pre-existing datasets to develop and validate novel clinical decision rules incorporating FAST and compare their performance to existing clinical decision rules (Aim 2). The proposed research and training plan will position Dr. Kornblith with cross-disciplinary skills to transition

to independence and submit a competitive R01 focused on refinement and validation of novel clinical decision rules integrating advanced computational methods applied to FAST for children after blunt abdominal trauma.