Low-Cost Early Prediction of Retinopathy of Prematurity Using Arterial Blood Gas Analysis: A Retrospective Cohort Study

For babies born prematurely, the world arrives too soon, and their bodies are still learning how to function outside the womb. Among the many challenges these fragile infants face, one of the most threatening is a condition called retinopathy of prematurity. This is a disease where the blood vessels in the back of the eye, the retina, do not grow properly. Instead of forming a healthy network, they can stop growing or grow in a chaotic, disorganized way, which can eventually lead to blindness. Because the eye is still developing, the timing of treatment is critical; doctors must find the problem early enough to intervene before permanent damage occurs. Currently, the standard way to screen for this is to take a camera and look directly inside the baby's eye. While effective, this process requires expensive equipment and highly trained specialists, and it often means checking many healthy babies just to find the few who are truly at risk.

Researchers have long known that the biggest risks for this eye disease are being born very early and having a very low birth weight. They also know that the amount of oxygen a baby receives plays a major role, as too much or too little can hurt the developing vessels. However, a new study suggests that doctors might be able to predict which babies are in danger much earlier and more simply, using data they already collect every day. By looking at the results of routine blood tests that measure how well a baby is breathing and how their blood is carrying oxygen, a team of scientists has built a computer model that can flag high-risk infants without needing to look inside their eyes first. This approach could help hospitals, especially those with fewer resources, decide which babies need the most urgent attention.

The study, conducted by a team of doctors and researchers in China, focused on over 500 premature infants born between 2022 and 2025. These babies were treated in neonatal intensive care units at three different hospitals. In these units, doctors routinely draw small amounts of blood from the babies to check their arterial blood gas levels. This test tells them how much oxygen and carbon dioxide is in the blood, how acidic or basic the blood is, and how well the body is using oxygen. The researchers gathered all this data, along with the babies' birth weights and how many weeks they were in the womb before being born. They then used a type of computer learning, where a program studies patterns in data to make predictions, to see if these blood test results could tell them which babies would develop the eye disease.

The team tested ten different computer algorithms to see which one worked best at spotting the risk. They found that the most successful model relied on a specific, manageable list of ten key pieces of information. These included the baby's gestational age and birth weight, plus nine specific numbers from the blood gas tests taken during the first six weeks of life. These numbers covered things like the pressure of carbon dioxide in the blood, the levels of different types of hemoglobin (the protein in blood that carries oxygen), and markers of how well the body was handling energy and stress. When the computer was trained on this specific set of ten facts, it became remarkably good at predicting the outcome. In the initial group of babies, the model achieved an AUC of 0.95. It also achieved an AUC of 0.91 for identifying healthy babies as low risk, meaning it rarely missed a case that needed attention.

To ensure this result was not just a lucky guess with one group of patients, the researchers tested their model on two other groups of babies from different hospitals. The model held up well in these new settings. In one group, it achieved an AUC of 0.83, and in the other, it achieved an AUC of 0.91. The study showed that adding more and more variables to the computer model did not make it better; in fact, the simpler model with just the ten most important facts performed the best. This suggests that the routine blood tests already being done in hospitals contain a hidden signal about the health of the baby's eyes that has been overlooked. The researchers noted that the model was particularly good at ruling out babies who were safe, which is a crucial feature for a screening tool. If a baby is flagged as low risk by this model, doctors can be confident that the baby is unlikely to develop the severe form of the disease.

The study also clarified what this tool is and what it is not. The researchers emphasize that this is a risk prediction system, not a replacement for the actual eye exam. A baby flagged as high risk by the model would not automatically receive treatment; instead, the result would tell doctors to look at that baby's eyes sooner and more often. This could be a vital difference for hospitals that do not have a specialist ophthalmologist available every day. By using the blood gas data that is already there, these hospitals could prioritize their limited screening resources for the infants who need them most. The model also worked for babies born at a slightly later stage than many other existing tools, which often only apply to the tiniest, most premature infants. This means the approach could help a wider range of babies, including those who are born a bit more mature but are still at risk.

The findings point to a deeper connection between how a baby's body handles oxygen and the development of their eyes. The blood tests revealed that babies who developed the eye disease had subtle differences in how their blood carried oxygen and how their bodies reacted to stress. For example, the levels of certain types of hemoglobin and the balance of acids in the blood were different in the at-risk group. These differences suggest that the body's struggle to manage oxygen and energy in the first few weeks of life is linked to the problems that later appear in the retina. By capturing these patterns, the computer model translates complex biological signals into a simple risk score.

While the results are promising, the researchers are careful to note that this is a retrospective study, meaning they looked back at data that had already been collected. The next step, which they are already planning, is to test this model in real-time as babies are being treated. They intend to make the tool available to other doctors and researchers so it can be validated in different parts of the world. If future tests confirm these results, this low-cost method could become a standard part of care for premature infants, helping to prevent blindness by ensuring that the right babies get the right eye exams at the right time. The work demonstrates that sometimes, the most powerful tools for saving sight are not new machines, but a smarter way of looking at the data we already have.