A research framework called AgentODE can infer the mathematical structure of a disease's dynamics using only summary statistics — no individual patient records required.
Researchers from arXiv introduced AgentODE to tackle a specific and stubborn problem: rare diseases rarely generate enough data to build reliable mechanistic models. The system uses a large language model to propose candidate ordinary differential equation (ODE) structures — essentially, equations that describe how a disease progresses over time. A second inference agent then iteratively refines the parameter distributions by running a diagnosis-update loop against population-level summary statistics. The team tested it on three benchmark problems and two clinical datasets, including one for recessive dystrophic epidermolysis bullosa, a rare skin disorder with only 231 observations across 46 patients.
The result cuts against an assumption baked into most medical AI: that more granular data always wins. In the RDEB experiments, models trained on individual-level data recovered implausible disease structures despite scoring better on raw predictive metrics. AgentODE, working only from aggregate statistics, found structures that were mechanistically coherent. That distinction matters — a model that fits the numbers but misrepresents the biology is a liability in clinical settings, not an asset.
AgentODE is an early sign that LLMs may have a practical role in scientific modeling, not just text generation — though the framework has only been tested on small, controlled benchmarks so far.