Bringing AI into the NICU: How algorithms may help infants’ eyes, health

When ophthalmologist Emily Cole, MD, steps into the neonatal intensive care unit (NICU) at the Children’s Hospital Colorado to evaluate an infant’s eyes for a disease called retinopathy of prematurity (ROP), it’s not uncommon for parents to decide to leave during the exam. 

“It can be heart-wrenching to watch,” says Cole, an assistant professor in the University of Colorado Anschutz Department of Ophthalmology. “These exams are difficult because these babies are very fragile, and we have to move them during their exam.”

But what if artificial intelligence could make the exam faster, more objective, and more precise? Going even further, what if AI could help detect not just ROP but other systemic diseases through the eyes, an emerging area known as oculomics? AI researcher Praveer Singh, PhD, also an assistant professor of ophthalmology, has been collaborating with investigators across the globe to develop AI-based algorithms designed to do exactly that.

With the goal of helping more patients, Singh and Cole have teamed up to further hone the capabilities of AI and understand the best ways for clinicians to apply these tools in the NICU.

“We’ve been developing all these algorithms, but the key is ensuring they have real clinical utility and can be effectively translated from bench to bedside,” says Singh, a faculty member in the Division of Artificial Medical Intelligence in Ophthalmology. “That’s what makes this collaboration so exciting.”

Why screening for ROP is challenging

One of the leading causes of childhood blindness, ROP is a disease that typically affects infants born early (before 31 weeks) or who weigh less than 3.3 pounds. It is caused by abnormal blood vessels in the retina, which is a layer of tissue near the back of the eyeball that detects light and transmits vision information from the eye to the brain.

“When fetuses are in the womb, the blood vessels grow to coat the inside of the eye. By 40 weeks, the retinal blood vessels have grown to cover the retina,” Cole says. “But for babies born early, their retinal blood vessels have not yet fully grown. These infants are then in the NICU, which tends to be an oxygen-rich environment, and that’s when abnormal vessels can grow in the retina, potentially leading to issues like bleeding, retinal detachment, and vision loss.”

An estimated 14,000 infants are diagnosed with ROP each year in the United States, according to the American Association for Pediatric Ophthalmology and Strabismus. Most of those cases are mild, with the disease not causing damage to the retina, but it’s estimated that each year, between 1,100 and 1,500 infants in the U.S. develop a more severe case of ROP that requires medical treatment.

The standard way to screen for ROP is to dilate an infant’s eyes and examine the retina for abnormal vessels and to see if the retinal blood vessels are abnormally shaped, which is performed by the pediatric ophthalmology team at Children’s Hospital Colorado.

“Taking a picture of a baby’s eye requires placing a camera on the surface of the eye with a numbing drop,” Cole says. “The exam can look scary because it is invasive, and it can cause the baby’s vitals to temporarily destabilize, such as their oxygen levels or heart rate decreasing, so we have to take breaks.”

From wiggles to scales

Not only are ROP eye exams difficult to perform, but there is also a shortage of providers who can screen and treat ROP, Cole explains. Treatments for ROP may include injecting medication into the eye, performing a laser procedure, or conducting surgery.

A key step for determining whether an infant needs treatment is assessing the retinal blood vessel tortuosity, meaning the crookedness of the blood vessels. When ROP is severe, the blood vessels can become wavy and wiggly.

“Typically, we describe how wiggly the vessels are — and how severe the disease is — by categorizing the condition as either ‘plus disease’, meaning very wiggly, ‘pre-plus disease,’ meaning kind of wiggly, or ‘not plus,’ meaning it is not wiggly,” Cole says. 

The issue is that not every clinician will categorize in the same way, Cole and Singh explain, which can lead to issues when patients are referred to another provider who may not agree with the original diagnosis.

“There is huge variability in terms of how one clinician is diagnosing ROP as compared to another clinician,” Singh says.

AI can help change that. Instead of using the three categories, clinicians can use an AI algorithm to scan an image of the retina — which was taken already as part of the ROP exam — and classify how tortuous the vessels are on a scale from one to nine, with nine representing the most severe disease. This is called a vascular severity score. Cole led research, published in 2022, that showed the potential of the AI tool to reduce the variability in diagnosing plus disease among patients with ROP. 

“The algorithm can predict disease severity using just one picture of the back of the eye. And for me, when I get that vascular severity score, that is a number that can help influence how I approach patient care,” Cole says. “This AI tool helps create a common language for all providers, and it hopefully will reduce our exam time, standardize diagnosis, and help us better predict which babies will need treatment.” 

Years in the making 

Singh has been working to advance AI technology to help screen for ROP for years. In 2022, while he was a postdoctoral research fellow at Harvard Medical School, he contributed to research that assessed the ability of an AI-based algorithm to diagnose ROP, specifically examining if the algorithm worked in external datasets from India, Mongolia, and Nepal.

“AI models often generalize poorly on external test sets, especially when imaging devices or patient demographics differ. Surprisingly, our algorithm performed pretty well,” he says. “A likely explanation is that, rather than operating directly on raw fundus images, which can differ substantially in intensity and pigmentation, we first segmented the retinal vasculature and then performed image analysis on those standardized vessel maps. This strategy improved robustness across external sites.” 

The research specifically used an AI-derived vascular severity score to identify infants who would develop treatment-requiring ROP. The study found that using the AI tools appeared to not only help identify high-risk infants, but it also could reduce the number of exams that low-risk infants endured.

“Extensive external validation has proven the effectiveness of this AI-based algorithm,” Singh says. “However, a key limitation is that many NICUs, especially in developing countries, cannot afford the expensive imaging device to capture the retinal images the algorithm requires.”

Looking for cost-effective alternatives, Cole and Singh, in collaboration with investigators from Oregon Health & Science University, tested the efficacy of smartphone-based telescreening for ROP as a more affordable alternative to expensive imaging cameras. They found that, despite lower image quality, a smartphone-based imaging device demonstrated a high probability of accurately detecting severe ROP. 

From computers to clinics

Given that the AI-based algorithm has been proven to be effective, the next step is determining how this tool can best be used in the clinic, Cole explains.

“On average, it takes 17 years for something that is in research to get into clinical practice,” says Cole, who has an interest in implementation science and aims to expedite that process. “From a speed and standardization standpoint, I wouldn’t be so psyched about putting this AI tool in the NICU if I didn’t really believe that this was helpful and efficient.”

Cole plans to interview a variety of stakeholders to get their input on this AI tool and how they think it should be implemented. Important stakeholders include parents, hospital administrators, data scientists, and health care providers like ophthalmologists, respiratory therapists, NICU nurses, and neonatologists. She will also look at measures of how clinicians adopt this technology through pilot trials in the NICU and how it affects their clinical workflows.

“There are a lot of possibilities for how this could be useful, and that’s what I want to explore,” she says, explaining the tool may not only be helpful in the NICU, but also for telemedicine, as a parent education tool, and as a handoff tool patients between providers.

Assessing for conditions beyond ROP

Using AI to scan images of an infant’s eyes is not only helpful in detecting ROP, Singh explains. It can also provide insight into other systemic health conditions.

Recently, Singh led research that examined whether retinal images obtained as part of the ROP exam may contain features associated with cardio-pulmonary diseases such as bronchopulmonary dysplasia (BPD) and pulmonary hypertension (PH), which are both leading causes of morbidity and mortality in premature infants. The research suggests that this retinal imaging-based AI tool can potentially predict the diagnosis of BPD and PH in premature infants, which may lead to infants being diagnosed earlier, thus avoiding the need for invasive diagnostic testing in the future.

“We found that we can detect these diseases way earlier by using AI instead of using invasive procedures like catheterization or echocardiogram,” Singh says. “I think there is tremendous potential for this tool, whether for early testing, or management, or even targeted intervention in high-risk infants.”

Cole adds, “What is unique is that, while he is building this algorithm and predictive model, I’m working in parallel to develop a workflow with neonatologists and NICU leaders at the University of Colorado Hospital and Children’s Hospital Colorado to determine where this predictive model would be most useful in their care.”

Singh is also contributing to the development of a foundational dataset exploring oculomics-based biomarkers for different systemic diseases in the neonatal population. At CU Anschutz, he’s helping create a clinical informatics and retinal imaging infrastructure focused on multi-morbidity, with hopes of expanding AI algorithms’ ability to detect a wider range of diseases through ocular image analysis.

Cole predicts one of the biggest challenges of bringing AI into the NICU will be getting clinicians on board with the change, which is why the implementation science approach is critical. 

“We’re going about this in a holistic way, where we are doing user testing at the same time as we are doing algorithm development. It’s important that we get the buy-in from clinicians and that they see this is worth it,” Cole says. “This tool can potentially help us predict who will need treatment, improve the speed of exams, standardize diagnosis, and improve access to care.”