How AI Could Change Clinical Trials for Ultra-Rare Disorders

Clinical trials were built for numbers. Rare disease often offers stories instead: a few dozen children across countries, symptoms that vary widely, a disease course that is poorly documented, and parents who know every small change but cannot always translate it into a regulatory endpoint.

This is the central problem of ultra-rare trials. The science may be promising, the community may be organized, and the need may be urgent. But the traditional trial machinery can struggle when there are too few patients for large randomized studies, too little natural-history data to know what would have happened without treatment, and too few validated measures to capture meaningful change.

AI could help in several practical ways. One is natural-history modeling. If researchers can combine registry data, clinical records, wearable signals, caregiver reports, imaging, and genetic information, models may help estimate how a disease tends to progress and which patients are likely to change over time. That can improve trial design because the study is no longer flying blind.

Another area is endpoint selection. Families often notice changes that standard scales miss: alertness, sleep, feeding, irritability, stamina, communication, comfort, participation. AI will not decide which outcomes matter, but it may help researchers analyze complex signals and identify measures that are sensitive enough to detect change without becoming meaningless.

Digital twins are part of this conversation. The term can sound more precise than the science currently allows, especially in rare disease. At its best, a digital twin is not a sci-fi copy of a person. It is a model that uses data to estimate what might have happened under different conditions. For small trials, that could help researchers compare treated patients against more informative predictions, though such approaches need careful regulatory and statistical scrutiny.

AI may also help with eligibility criteria. Rare disease trials can accidentally exclude the very people they hope to help if criteria are too narrow, too burdensome, or based on assumptions that do not match real families. Better modeling may help sponsors understand which criteria are essential for safety and which simply make recruitment harder.

The risk is that complexity becomes camouflage. A trial can be powered by AI and still ask the wrong question. A model can look sophisticated and still be trained on biased or incomplete data. A digital comparator can be useful and still fail if regulators, clinicians, families, and statisticians cannot understand the assumptions behind it.

For families, the goal is not easier approval. The goal is better evidence gathered in a way that respects small populations and real lives. A good ultra-rare trial should be rigorous, but it should also be humane. It should ask whether a therapy changes something meaningful, whether the risk is justified, and whether the design gives the community a fair chance to learn.

AI can help with that, but only if it is treated as a tool for discipline rather than a shortcut around uncertainty.