Virtual control arms for paediatric myopia trials: external validation of axial-elongation models

The human eye continues to grow throughout childhood, much like the rest of the body. In many children, this growth does not stop when it should, causing the eye to become too long. This condition, known as myopia or nearsightedness, makes distant objects appear blurry. For decades, the standard way to test whether a new treatment could slow this growth was to run a clinical trial. In such a trial, some children would receive the new treatment while others received a standard pair of glasses with no special features, serving as a control group to show what would happen without intervention. However, as effective treatments for slowing eye growth have become more common, it has grown increasingly difficult and ethically questionable to ask families to enroll their children in a study where they might receive no active treatment at all. Parents are reluctant to let their children go without help when better options exist, and children who are getting worse quickly often drop out of the study, skewing the results. This creates a dilemma for researchers: they need to know what happens without treatment to prove a new drug or lens works, but they cannot easily find children willing to be untreated.

To solve this, researchers have turned to mathematics. Instead of relying on a group of untreated children, they can use computer models to predict how a child's eye would grow if they received no treatment. These models act as a "virtual control arm," providing a statistical estimate of the untreated path based on factors like the child's age, current vision, and background. The challenge, however, is knowing if these predictions are actually accurate. If the model is wrong, the entire trial could be misleading. A team of researchers set out to test five different published models to see if they could correctly predict the eye growth of real children who were not receiving any special treatment. They gathered data from 242 children of Chinese, Vietnamese, and Indian ancestry who were wearing standard glasses. The researchers measured the length of these children's eyes at the start of the study and again after about six and twelve months. They then compared the actual growth of these real eyes against the growth predicted by the five different mathematical models.

The study found that the most recent models, which take into account a child's specific ethnic background, were remarkably accurate for children of East Asian descent. For these children, the models predicted the average amount of eye growth over six months with a tiny margin of error, differing from the real measurements by less than the width of a human hair. This suggests that for this specific group, a virtual control arm could reliably replace a group of untreated children in a clinical trial. However, the researchers also discovered a significant limitation: the models were not perfect at predicting how much variation exists between individual children. While the models got the average growth rate right, they underestimated the range of differences seen in real life. The models suggested that most children would grow at a very similar pace, but in reality, some children grew much faster or slower than the average. This means the models are excellent tools for predicting what happens to a large group of children, but they cannot accurately predict the future growth of a single specific child.

The situation was more complicated for children of South Asian descent, specifically those of Indian ancestry. None of the existing models had a specific category for this group. When the researchers applied the models meant for East Asian or European children to the Indian children in the study, the predictions were off. The Indian children's eyes grew at a rate that fell somewhere between the rates predicted for the other two groups, closer to the East Asian rate but not matching it exactly. This indicates that while the current tools are useful, they are incomplete. They work well for East Asian populations but leave South Asian children without a reliable prediction method. The researchers concluded that for the models to be truly useful for everyone, new data specific to South Asian children must be collected and added to the system.

Another key finding was that the timing of measurements matters more than previously thought. The models are designed to predict growth over a specific period, such as exactly six months. In the real world, children's check-up appointments rarely happen on the exact calendar date. The study showed that if researchers simply assumed a visit happened at six months when it actually happened at six and a half months, the prediction would be slightly wrong. By using the exact number of days that passed between visits, the researchers were able to make their comparisons much more precise. This attention to detail is crucial because the difference between a successful treatment and a failed one can be very small, often just a fraction of a millimeter in eye length.

The researchers also developed a free, open-source software tool that brings all these different models together in one place. This tool allows other scientists to run the same calculations on their own data without needing to write complex code or share sensitive patient information with a central server. The software runs entirely on a user's own computer, ensuring privacy. It does not claim to predict the future of an individual child's eyes with certainty. Instead, it provides a scientifically defensible estimate of what a group of untreated children would look like, which can serve as a reference point for new treatments. The study confirms that for East Asian children, these virtual controls are a viable alternative to traditional untreated groups, potentially making clinical trials easier to run and more ethical. For other groups, particularly South Asian children, the work highlights a gap in current knowledge that needs to be filled. The ultimate goal is to make the use of untreated control groups in children the exception rather than the rule, relying instead on validated, transparent mathematical models to protect the well-being of young patients while advancing eye care.