2224. Performance of an Endotracheal Tube Detection AI Algorithm on Portable Chest Radiographs in Children 10-18 Years
Authors* Denotes Presenting Author
Shannon Sullivan *;
University Hospitals of Cleveland
Kaustav Bera;
University Hospitals of Cleveland
Amit Gupta;
University Hospitals of Cleveland
Sheila Berlin;
University Hospitals of Cleveland
Objective:
The incidence of endotracheal tube (ETT) malposition in children is reported at 15% to 30%. FDA-approved artificial intelligence (AI) tools for the detection and localization of ETT are available for use in adults. As these AI tools have not been validated in children, we aimed to assess the performance of an on-device algorithm for ETT identification and ETT tip and carina localization on portable chest radiographs (CXR) of pediatric patients aged 10 to 18 years.
Materials and Methods:
This retrospective study included 535 consecutive CXR in 10 to 18-year-old children in our pediatric Level 1 Trauma ED and ICU collected from 2016 - 2022; CXR were included irrespective of image quality. 286 CXR had an ETT present; 249 CXR had no ETT. Two radiologists-in-training with similar experience levels performed ground truth annotations for ETT and carina locations; they also measured the ETT-to-carina distance. Next, an FDA-approved AI tool for ETT tip and carina detection and ETT tip-to-carina measurement was applied. The ground-truthers then assessed whether the AI tool appropriately detected the ETT tip and carina. AI tool detection was scored as appropriate if = 7mm from the ground truth location. ETT positioning was classified based on distance from carina: endobronchial (< 0 cm), too low (< 1.5 cm) or too high (>6 cm). Disagreements were arbitrated by a senior pediatric radiologist. Overlying foreign bodies (FB) in the region of the ETT and carina such as tubes and lines were recorded. Image quality scored as excellent, good, average or poor. Fisher’s exact test assessed ETT detection performance with subgroup analysis between ages and gender. Cohen’s Kappa assessed agreement in ETT positioning.
Results:
AI had sensitivity, specificity and likelihood ratio of 0.99, 0.99, 246 (<em>p</em> < 0.01) in ETT detection. One false positive occurred in a CXR with a tracheostomy tube. Four false negatives occurred in CXRs with scoliosis curves of 27, 40, 50 and 56 degrees. AI correctly identified 98% (47/48) of all malpositioned ETT: 2/2 endobronchial, 17/17 too low and 28/29 too high ETT. Overlying FB were present on 66% of CXR with > 2 FB present in 26%. Image quality scores were average or poor in 51%. Fisher’s exact test for subgroup analysis between age groups in years (10 - 13; 13 - 16; 16 - 18) and gender showed no statistically significant differences in ETT detection performance. The appropriateness of AI localization was 95.8% for the ETT tip and 94.7% for the carina. There was almost perfect agreement between ground truth and AI classification of ETT positioning with Cohen’s Kappa = 0.853 (95% CI = 0.78-0.93) despite frequent poor image quality and multiple overlying FB.
Conclusion:
An FDA-approved AI algorithm for adults shows similar performance in ETT identification and localization on pediatric CXR despite frequent poor image quality and multiple overlying FB. Nearly all critically-malpositioned ETT were correctly identified. Use of a robust on-device algorithm can flag critically-malpositioned ETT for prompt review and management in this complex pediatric ICU and ED population.