Machine Learning Algorithm Can Diagnose FBA at Level Comparable to Pediatric Radiologists 

by Linda Kossoff •

CLINICAL QUESTION

What is the potential of machine learning to improve the diagnostic capabilities of chest radiographs in cases of pediatric foreign body aspiration (FBA)?

BOTTOM LINE

Chest radiograph analysis augmented with machine learning can diagnose FBA in pediatric patients at a level similar to that of a pediatric radiologist’s read.

BACKGROUND: Foreign body aspiration is a life-threatening condition seen primarily in young children. Prompt diagnosis of FBA is critical but challenging. Chest radiographs alone are insufficient. Bronchoscopy is effective but carries risks and is not appropriate for everyone. Machine learning, a subset of artificial intelligence (AI), may help in diagnosing FBA.

STUDY DESIGN: Retrospective study

SETTING: Division of Pediatric Radiology, Department of Radiology and Biomedical Imaging, University of California, San Francisco

SYNOPSIS: Researchers performed a retrospective chart review of 566 pediatric patients (352 males; mean age 2.89 years) presenting with a potential diagnosis of FBA between 2010 and 2020. A total of 536 chest radiographs were extracted from charts, 1,128 from the National Institutes of Health’s ChestX-ray8 database, and 24 from Google Images, for a total of 1,688 radiographs. Of these, there were 1,480 negative findings and 208 positive findings of FBA. All radiographs were uploaded to Google AutoML Vision, which was used to develop an algorithm based on machine learning-driven image classification. The sensitivity and specificity of the radiologist interpretation were shown to be 50.6% and 88.7%, respectively. Comparatively, the sensi-tivity and specificity of the algorithm were shown to be 66.7% and 95.3%, respectively. The precision and recall of the algorithm were both 91.8%, with an average precision of 98.3%. The authors note that the algorithm utilized only frozen images of frontal view chest radiographs and did not take into consideration any other information that might improve its accuracy. Study limitations included the inherent circumstances of machine learning, which is confined to the available database.

CITATION: Truong B, Zapala M, Kammen B, et al. Automated detection of pediatric foreign body aspiration from chest X-rays using machine learning. Laryngoscope [published online ahead of print February 17, 2024]. doi:10.1002/lary.31338.