A senior clinician can walk into a room, glance at a patient, and sense that something is wrong before a single test comes back. Medicine has a name for that instant read: clinical gestalt. It is not a guess, and it is not magic. It is expert pattern recognition, built on thousands of real cases. Over a career, the clinician has internalised a dense web of clinical knowledge. The hard part of teaching medicine is that you cannot simply hand that web to a student. They have to build it, case by case, over years.
A knowledge graph is a synthetic version of that web: the same dense network of clinical connections, built deliberately so it can sit behind Gestalt and be practised against at scale.
A medical knowledge graph is a structured network of clinical concepts, findings, diseases, investigations, medications, and management, joined by the relationships that clinicians actually use to think. The Gestalt knowledge graph is built on clinical terminology such as SNOMED, together with trusted clinical knowledge partnerships and published clinical references. Every AI interaction in Gestalt is constrained by it.
This post is about what that graph is, what it lets students do, and why it is the part of Gestalt that we believe will set the standard for the future of medical education. It builds on two earlier posts: one on why a general chatbot is the wrong thing to lean on while you are still learning, and one on why medicine needs grounded probability rather than ungrounded text.
What is a medical knowledge graph?
Think of it as a map of clinical reasoning. The points on the map are the things clinicians reason about: a symptom such as pallor, a disease such as anaemia, an investigation such as a blood film, a medication and its cautions. The lines between them are the relationships that matter: this finding points towards that disease, this disease calls for that investigation, this medication treats that condition but not in this patient. A knowledge graph holds both the points and the lines, in a form a computer can reason over.
That is very different from what a general-purpose language model has. A language model has learned how words tend to follow other words. It knows that chest pain and heart attack appear together often in writing, because they do. What it does not have is a verified, sourced relationship saying that a particular pattern of chest pain raises the probability of a particular condition by a particular amount. It has the language of medicine without the structure underneath it. The graph is that structure. As we argued in an earlier post, clinical reasoning is the skill medical school is really trying to build, and you cannot practise reasoning against a system that only knows how the words sound.
Why does Gestalt have one when most AI products don't?
Grounding medical AI in a structured knowledge graph has long been argued to be the soundest way to build it, and that view is fast becoming the mainstream position rather than a niche one. A 2025 review of knowledge graphs in healthcare describes pairing structured medical knowledge with language models as a leading direction for safer clinical AI, and peer-reviewed studies show that grounding a language model in a medical knowledge graph improves its diagnostic reasoning. Most of that published work is about clinical decision support rather than education, and the demands of teaching are not identical to the demands of the clinic. But we believe medical education deserves the same robustness as medicine itself, which is why we built Gestalt on the same foundation.
It is also hard, slow work. The quickest way to ship any AI tool is to take a frontier language model, write a clever prompt, and let it talk. That is a wrapper. It looks impressive in a demo, but it inherits every weakness of the model underneath, including its habit of inventing facts with complete confidence. It is also exposed to any change or restriction the provider imposes later. That is not something an education provider can or should rely on.
A knowledge graph is the opposite of a shortcut. It means deciding what counts as a clinical concept, encoding how concepts relate, and attaching a source to each claim. You will hear this kind of approach described in a few ways: a knowledge graph, a clinical ontology, or knowledge-graph-augmented generation. What it is not is a handful of documents dropped into a model's context window in the hope it reads them properly. It is a high-density network of verified connections that sits underneath every interaction and bounds what the AI is allowed to say.
This is the foundation the whole product is built on. If you want to understand what makes Gestalt different from anything else a student could open in a browser, it is the knowledge graph, and because one structured body of clinical knowledge can power every mode of learning, it is also what lets the approach scale.
What's actually inside the Gestalt knowledge graph?
At its core is a set of relationships that mirror how clinicians organise what they know. Findings point to diseases, diseases call for investigations, and investigations lead to management. Diseases connect to their related conditions, complications, and risk factors. Medications connect to their indications, contraindications, side effects, and interactions.
Those concepts are anchored to clinical terminology such as SNOMED, the international standard for coding clinical ideas. Trusted regional knowledge partnerships supply locally relevant medicines and treatment information. Published clinical references supply the rest. The relationships are not simply present or absent; they carry clinical weight, capturing how strongly a given finding points towards or away from a given diagnosis. That weighting is what lets the system reason rather than merely look things up.
How the graph is traversed, the precise schema of its relationship types, and the way the language model and the graph hand off at run time are our secret sauce, and they stay internal. What you can trust is the shape: a structured, weighted map of medicine, not a wrapper with a prompt and a few documents.
What does grounding AI in a knowledge graph let students do that an LLM wrapper can't?
Start with reasoning. Because a differential is computed against verified relationships between findings and diseases rather than generated as plausible-sounding text, the reasoning a student sees is real reasoning they can follow and question. The system weighs competing diagnoses the way a clinician does, and a student can interrogate why one moved up the list, because there is an actual structure underneath the answer rather than a turn of phrase.
Then there is connectedness. Because findings, diseases, investigations, and management are all linked in one graph, the medicine a student meets is consistent wherever they meet it, and they can follow the threads between concepts rather than memorising isolated facts. Land on a finding like pallor and the system can show what it might point to, the other findings you would expect alongside it, and the conditions and investigations it connects to. That is what turns noticing that a patient looks pale into beginning to understand why it matters.
There is also local relevance. Medicine is not the same everywhere: guidelines, available medicines, and accepted practice differ from country to country. Because the clinical knowledge lives in the graph rather than being baked into the model, Gestalt can be localised without rebuilding the system around it.
That is what lets the same product carry locally relevant medicine wherever a student is training, instead of defaulting to whatever jurisdiction happened to dominate the model's training data. It is also what makes the approach scalable: the clinical reasoning is built once, and only the local knowledge changes.
Who oversees the medical knowledge, and how do we keep it honest?
The medical knowledge is designed and governed by a clinician. Gestalt's Chief Science Officer, Dr Nick Wright, built the clinical knowledge architecture that sits behind the product. He is an experienced general practitioner with fifteen years in practice, a published researcher whose work has appeared in the Nature portfolio of journals, including a study that ranks in the top 0.1% of all research by attention, and a former teacher of clinical reasoning at the University of Otago.
The AI works downstream of that expertise: the language model is only ever allowed to express what the clinical knowledge supports. The medicine comes first, and the model serves it.
That same order, knowledge first and model second, is what keeps it honest. The knowledge in the graph carries its sources, so the clinical content behind an interaction is tied to where it came from rather than taken on trust, and the model is held to what that knowledge supports rather than left free to improvise. This is not just our preference: recent research finds that grounding a language model in a structured clinical knowledge graph measurably reduces the confident fabrications that general models are prone to.
Why is this more than a clever interface?
Plug an off-the-shelf avatar into a general language model, add a prompt, and you can get something that looks like a working patient simulator. There are tools that do exactly that, and for a quick demonstration it can be enough. But for teaching the practice of medicine, for doing right by your students, and for where this technology is heading in medical education, that is not enough. A tool that simply passes a student's questions to a general model borrows the model's fluency without adding any accountability. It can sound convincing whether or not it is right, and it will state a wrong answer with the same confidence as a correct one. The difference is whether there is verified, structured medicine underneath the conversation, or nothing at all.
Gestalt is built to be simple to use. A student should be able to open it and practise without thinking about what runs underneath, and that is deliberate: the hard engineering stays out of the way. But for those serious about the future of medical education, what runs underneath is the whole point. The knowledge graph is what makes Gestalt different, and it is the foundation everything we build next will stand on.