Researchers can now train clinical AI on 23 patients instead of thousands

A new method generates synthetic patients from tiny real cohorts, preserving the statistical and biological structure of the original group. This means hospitals studying rare diseases or pregnancy complications can now build predictive models from datasets too small to train on before.

Clinical trials in maternal health, rare diseases, and early-phase studies have always faced a hard constraint: too few patients to train reliable AI models, but too expensive and slow to recruit more. This technique breaks that constraint by generating synthetic patients that behave identically to real ones across multiple validation tests, including mechanistic models of biological systems. The immediate effect is that researchers can now augment tiny cohorts without waiting years for enrollment. The longer effect depends on whether hospitals and trial sponsors actually adopt this — if they do, rare-disease research accelerates and the cost of early-phase modeling drops.