252 - A robust deep learning model that localizes the endotracheal tube tip and mid-tracheal position to evaluate the adequacy of the endotracheal tube tip position in critically ill neonatal patients.

Poster Number: 252
Publication Number: 252.3030

Background: The care provided in the neonatal intensive care unit is holistic. Among them, the most important treatment is respiratory therapy. The proper placement of the endotracheal tube within the patient's airway is a crucial step. If the endotracheal tube is not properly positioned, it can result in serious complications. Extensive research has been conducted for estimating and confirming the appropriate insertion of endotracheal tubes. It is essential to visually confirm the placement of endotracheal tubes by capturing portable AP CXR after intubation. Notably, artificial intelligence algorithms have demonstrated strong performance in assessing ETT placement on chest radiographs. However, most studies targeted adult patients and used single-center data or public data, and there were no studies on neonatal critically ill patients using multicenter data.

Objective: This study involved the development of a robust deep learning model for localizing ETT tip and the mid-trachea position, a critical landmark for ETT placement, in NICU CXR.

Design/Methods: Of the 10 neonatal intensive care units in Korea, 1,000 images were selected through random sampling from 6057 images obtained from 9 institutions and used for training, internal validation, and internal testing, and the 100 images from 1 institution were used for external validation. In neonatal CXR from the NICU, we used a CNN-based object detection model to detect the landmarks T1 and T2, which are used to assess the appropriateness of endotracheal tube placement, including the tip of the endotracheal tube. Predicted endotracheal tube tips were assessed for error using Euclidean distance (unit: mm), while predicted T1 and T2 regions were evaluated using average precision (AP).

Results: The mean absolute error of the model on the internal test set was 1.58 mm ± 1.16 for the ETT tip position prediction, 1.28 mm ± 0.98 on the external test set. The average precision of the model on the internal test set was 0.969 for mid tracheal position (T1, T2), 0.95 on the external test set. The difference between internal and external results is not statistically significant.

Conclusion(s): we have developed a deep learning model that demonstrates strong performance in both internal and external settings. This can be attributed to the learning from a diverse set of neonatal chest X-rays collected from nine neonatal intensive care unit (NICU) and the high-quality annotations achieved. The use of the model can potentially enhance the evaluation of the appropriateness of ETT tip positions, contributing to the improvement of clinical efficiency and reduction of human errors.