Prediction of Endotracheal Tube Depth by Using Deep Convolutional Neural Networks

Chang Gung Medical Foundation

Status: Completed

Conditions: Machine Learning; Intubation

Treatments: Diagnostic Test: Deep convolutional neural networks analysis

Study type: Observational

Funder types: Other

Identifiers

NCT05085743

202002007B0

Details and patient eligibility

About

Malposition of an endotracheal tube (ETT) may lead to a great disaster. Developing a handy way to predict the proper depth of ETT fixation is in need. Deep convolutional neural networks (DCNNs) are proven to perform well on chest radiographs analysis. The investigators hypothesize that DCNNs can also evaluate pre-intubation chest radiographs to predict suitable ETT depth and no related studies are found. The authors evaluated the ability of DCNNs to analyze pre-intubation chest radiographs along with patients' data to predict the proper depth of ETT fixation before intubation.

Full description

This was a retrospective, IRB-approved study using chest radiographs images obtained from Picture Archive and Communication System (PACS) at Chang Gung Memorial Hospital, Linkou branch, Taiwan.

A total of 595 de-identified patients' chest radiographs was obtained for this study. The inclusion criteria for this study are patients 18 years or older who were orotracheal intubated within November 2019 to October 2020 and had taken chest radiographs before and immediately after the intubation (<24 hours). Both pre-intubation and post-intubation chest radiographs of a same patient were obtained. Clinical data including age, sex, body height, body weight, depth of ETT fixation were also recorded. All ETT tip to carina distance was manually measured by a same anesthesiologist from post-intubation films and documented. Lip to carina length of each patient can be calculated by adding ETT fixation depth and ETT tip to carina distance.

Pre-intubation chest radiographs (n=595) along with clinical data including age, sex, body height, body weight, and measured lip to carina length are collected for model building. For this study, 476/595 (80%) of those were used for training and 119/595 (20%) for validation randomly selected by AI model. In training process, images and related clinical data along with the measured lip to carina length are fed into and used to fit out AI model. Then, in validation process, the investigators evaluate the model accuracy and efficacy of predicting the lip to carina length with images and clinical data of those unforeseen cases.

Enrollment: 595 patients

Sex: All

Ages: 18+ years old

Volunteers: No Healthy Volunteers

Inclusion criteria

• 18 years or older
• orotracheal intubated within November 2019 to October 2020
• had taken chest radiographs before and within 24hr after intubation

Exclusion criteria

• Bad chest radiographs quality that patients' carina can not be recognized
• Patient with bronchial insertions found in post-intubation films
• Nasal intubation

Trial design

595 participants in 2 patient groups

Training

Description:

Images and related clinical data along with the measured lip to carina length of the training group are fed into and used to fit out deep convolutional neural networks model.

Treatment:

Diagnostic Test: Deep convolutional neural networks analysis

Validation

Description:

We evaluate the model accuracy and efficacy of predicting the lip to carina length with images and clinical data of those unforeseen cases in the validation group.

Treatment:

Diagnostic Test: Deep convolutional neural networks analysis