Contact: Emily Greendonner
MADISON, Wis. – Researchers at the University of Wisconsin School of Medicine and Public Health are developing an artificial intelligence-based “digital mentor” to support the training of clinicians who help newborns in the neonatal intensive care unit breathe.
When babies are born too early, they can have difficulty breathing because their lungs are not fully developed. Some infants need doctors to place a breathing tube in their windpipe and connect it to a machine that helps them breathe. This procedure is called intubation, and it is one of the highest-risk procedures performed in the NICU.
During intubation, clinicians use a tool called a laryngoscope to visualize a newborn’s airway and place a breathing tube.
There are two main approaches. With direct laryngoscopy, the clinician peers directly into the infant’s mouth and airway while using the device to see the vocal cords. The second approach, called video laryngoscopy, is becoming more common as technology has advanced. The video laryngoscope has a small camera that displays the airway on a screen, allowing a team approach to guided intubation.
Video laryngoscopy is increasingly used, although either technique may be chosen depending on the clinical circumstances, equipment, and clinician experience, according to Dr. Ryan McAdams, professor of pediatrics, University of Wisconsin School of Medicine and Public Health.
For premature and critically ill infants, the margin for error is small. Prolonged or repeated intubation attempts can cause low oxygen levels, low heart rate, airway trauma and other complications. Learning the important skill of intubation can require up to 100 supervised attempts, and opportunities for doctors to practice can be limited, he said.
“We realized we had a clinical problem: How can we teach this skill as safely and quickly as possible?” said McAdams, who is also a UW Health Kids neonatologist at American Family Children’s Hospital’s Level IV NICU and at UnityPoint-Health Meriter Hospital’s Level III NICU. “We turned to our colleagues on campus with expertise in deep learning and AI to help leverage new technology to develop a solution.”
McAdams, along with his colleague Dr. Patrick Peebles, assistant professor of pediatrics, collaborated with Yin Li, associate professor of biostatistics and medical informatics, at the school, along with Abrar Majeedi, a recent UW–Madison doctoral graduate in the Department of Biostatistics and Medical Informatics whose research focused on deep learning and health AI. The team adapted two established AI systems, called YOLOv8 and I3D, to analyze neonatal intubation videos.
Video laryngoscopies were collected at the hospitals and de-identified to protect patient privacy for research purposes. McAdams and Peebles reviewed more than 100 videos, totaling nearly 300,000 frames, to identify key airway anatomy, such as the space between the vocal cords, referred to as the glottic opening, and to classify the depth of laryngoscope blade insertion. The goal was to develop technology that helps clinicians recognize what they are seeing, understand blade position and receive more objective feedback during training.
“We compare this to a backup camera on a car. It provides visual cues to improve situational awareness by circling and highlighting key areas, but the technology does not replace the person performing the task,” McAdams said. “Instead, it serves as a strong helper or a good coach for a learner. This feedback can be helpful for experienced clinicians as well.”
UW School of Medicine and Public Health researchers published two papers on their AI work in the Journal of Perinatology in the past two years. The first paper showed that the AI model could identify the vocal cord region. Its performance was comparable to or better than that of novice and intermediate clinicians, although experienced clinicians identified the anatomy more quickly, according to Peebles.
The UW team was among the first to study how deep-learning technology could support neonatal video laryngoscopy using recordings from actual clinical care, he said.
“The ultimate goal is to integrate this AI technology into every video laryngoscope for all clinicians in the future, but several steps must be taken before that can happen,” said Peebles, who is also a UW Health Kids neonatologist at American Family Children’s Hospital and Meriter. “This AI model has the potential to help, especially serving medical teams in more rural or developing areas that don’t perform it often and thus may be less familiar with the procedure.”
If the technology is successfully refined and validated, it could expand access to procedural coaching and potentially improve the safety and success of neonatal intubation, he said.
The second paper examined blade depth and positioning. While the research team needs more data before drawing definitive conclusions, the findings showed potential for more precise coaching, according to Peebles.
For now, the work remains in the model-development and evaluation stage, with promising potential for application in procedural training and quality improvement. Researchers are taking a staged approach to evaluating how computer vision could support neonatal procedural training and, eventually, bedside decision support, according to McAdams.
“Establishing an airway can be lifesaving for critically ill babies,” he said. “We have an outstanding team looking at how new technology might make the procedure safer and more effective.”
