An AI system for wound care triage tries to solve medicine's LLM problem by keeping the model on a short leash.
Researchers have published the architecture and validation results for Medi-Gemma, a clinical decision support system designed for wound pathology triage and workflow automation. Rather than letting a large language model handle all queries, the system routes requests through a tiered pipeline: a component called the IntentRouter sends analytical questions to deterministic code executed via a PandasQueryEngine, while patient-specific queries pass through a retrieval-augmented generation engine. A "Ground Truth Injection Module" pulls verified numeric data directly from structured patient records and embeds it into the prompt before the model ever generates a response — effectively forcing the AI to work from confirmed facts rather than infer them.
The approach matters because hallucination in a clinical context is not a user-experience flaw — it is a patient safety risk. Most LLM deployments in healthcare have stumbled on exactly the problems Medi-Gemma targets: models that drift from structured data, fabricate lab values, or omit critical information when retrieving from vector stores. By separating data orchestration from generative inference and adding a deterministic ProtocolManager that maps clinical language to fixed evidence-based risk pathways, the system's authors claim they eliminated semantic context drift and improved factual adherence to backend records.
The paper's honest contribution is less the model itself and more the architectural pattern — a blueprint for constraining LLMs in high-stakes environments by making the deterministic layer sovereign and the generative layer subordinate.
Whether this holds outside a research validation setting, in the hands of real clinical staff with messy real-world EMR data, is a question the paper does not answer.