Scientists have developed an artificial intelligence-powered model that could help identify premature babies at risk for pulmonary hypertension (PH) and bronchopulmonary dysplasia (BPD) using non-invasive eye photos taken during standard screening.
Dr. Jayashree Kalpathy Kramer, professor of ophthalmology at the University of Colorado Anschutz and senior author of the study, said: “These findings suggest that information about the health of a baby’s lungs and heart may already be present in these images routinely collected during neonatal care. “Early detection can make a meaningful difference in outcomes and treatment plans,” said Dr. Jayashree Kalpathy Kramer, professor of ophthalmology at the University of Colorado Anschutz and senior author of the study, in a university press release.
The study, “Deep learning-based prediction of cardiopulmonary disease in retinal images of premature infants,” was published in the journal JAMA Ophthalmology.
PH is a complication of BPD and can increase risk of severe disease
BPD is a chronic lung disease most common in premature infants. It damages the airways (bronchial tubes) and prevents the development of small air sacs in the lungs. PH can be a complication of borderline personality disorder and increases the risk of serious illness.
Diagnosing BPD and PH early in preterm infants can be difficult because BPD prediction tools and PH diagnosis rely on tests and techniques that may not be sufficiently sensitive or are resource and time intensive.
“Obtaining surrogate biomarkers that correlate with early stages of BPD and PH through non-invasive techniques may avoid the need for more invasive testing and allow neonatologists to provide more individualized and customized care,” the researchers wrote.
With this in mind, researchers developed an AI-powered model to analyze retinal images taken during routine examinations for retinopathy of prematurity (ROP). ROP affects the retina (the light-sensing tissue at the back of the eye) in premature babies.
The analysis included a total of 493 premature infants with retinal images collected between June 2015 and April 2020 as part of the national Imaging and Informatics in Retinopathy of Prematurity (i-ROP) study.
“One of the challenges in realizing the potential of many ophthalmic algorithms is that imaging the back of the eye is not (yet) part of the usual treatment pathway for many patient populations,” said study co-author Peter Campbell, MD, professor of ophthalmology at Oregon Health & Science University. “For more and more NICUs [neonatal intensive care units]Imaging is part of the treatment pathway for ROP, meaning the barriers to implementing such technology are significantly lower. ”
Improved performance of composite models
The images were taken within 34 weeks after menstruation (from the first day of the mother’s last menstrual period), and therefore before BPD and PH are typically diagnosed.
To assess whether photos of eyes could indicate new problems with the lungs or heart, researchers trained an AI model using three approaches. Retinal imaging only, demographic and clinical risk factors only, and a combined model that integrates both.
In this study, BPD was defined as a baby requiring supplemental oxygen at 36 weeks postmenstrual age, and PH was defined based on an echocardiogram at 34 weeks postmenstrual age and specific pulmonary artery pressure criteria.
The team then tested how accurately the AI model could predict which infants would meet these criteria using a standard measure called the receiver operating characteristic (ROC) area under the curve (AUC). AUC values range from 0 to 1, with 0.5 indicating random chance and higher values indicating a better ability to distinguish between babies with this condition and those without.
Retinal images obtained during ROP screening may be used to predict the diagnosis of BPD and PH in preterm infants, leading to early diagnosis and potentially avoiding the need for invasive diagnostic tests in the future.
Overall, the performance of the combined model has improved. For PH, the combined model achieved an AUC of 0.91. This was the same as for images alone, but was better than 0.68 for demographic risk factors alone.
For BPD, the AUC of the combined model reached 0.82, while the AUC of the demographics-only and images-only models was 0.72.
Importantly, the results remained even when the researchers removed images that showed signs of ROP, suggesting that the model was not simply “reading” the severity of the eye disease.
“Artificial intelligence allows us to detect subtle patterns in retinal images that are invisible to the human eye,” said Praveer Singh, Ph.D., assistant professor of ophthalmology at the University of Colorado Anschutz and lead author of the study. “This opens up the possibility of using simple photographs to gain insight into a premature baby’s overall health.”
The researchers highlighted the model’s potential to aid in diagnosis.
“Overall, these findings concluded that “retinal images obtained during ROP screening may be used to predict the diagnosis of BPD and PH in preterm infants, leading to early diagnosis and potentially avoiding the need for invasive diagnostic tests in the future.”
