Questions from the people who run care delivery.
A clinical context layer is a persistent, continuously maintained model of how a patient's facts relate to one another: records, observations, devices and clinician concerns. Applications draw the context they need from it. Clinical Context AI (IMI) is specifying one. It proposes, and a clinician decides.
The electronic health record is a system of record: it preserves what was decided and documented. A clinical context layer holds how those facts, and signals outside the record, relate to one another over time. It does not replace the record. It draws on it and can feed insight back through alerts and decision support.
No. Clinical Context AI (IMI) is specified, with a prototype next, and is not in clinical use. The patient examples in our articles are synthetic and say so. If you would like to discuss the architecture, talk to the founders.
No. The system proposes and a clinician decides. It never asserts a clinical fact about a patient on its own, and no clinical action is automated. Its job is to keep the relevant context connected and visible so a clinician can judge it.
Questions from digital health, remote monitoring and device teams.
In the architecture described in the whitepaper, event sources such as devices, applications and systems of record send events to the layer. Each event is interpreted against the existing context, and the model is updated. This describes the specified design; it is not a description of a deployed integration.
Yes, by design. The context layer sits between the systems being observed and the applications that reason about them. Each application builds a scoped context window from the shared model for its own task, so contextual understanding is built once and reused instead of rebuilt in every application.
No. Monitoring and analytics still produce the signals. A clinical context layer holds the relationships between those signals and the rest of what is known about the patient, so that a slow drift in several measurements can be seen together instead of one threshold at a time.
Stage, thesis and what is, and is not, built.
Clinical Context AI (IMI) is an early-stage team, founded in 2026. The context layer is specified, with a prototype next, and is not in clinical use. The investor page describes the thesis, architecture, team and where things stand, and the deck is available on request.
Clinical Context AI was founded by Keith Deutsch, co-founder and CTO, and Ratnadeep Bhattacharjee, co-founder. Keith has served as CTO and Chief Architect at Motive Medical Intelligence and VP Engineering at Cancer Commons, and at GE originated the Active Context Graph architecture for the Predix platform. Ratnadeep previously co-founded and led TechVariable, an AI and analytics company serving digital health clients.
How a context layer relates to context windows, retrieval and agents.
A context window is the information given to a model for one reasoning task, assembled for that inference and then discarded. A context layer is a persistent model of the relationships among facts, maintained independently of any inference, from which many context windows can be drawn. One is temporary; the other is cumulative.
Retrieval finds information that is semantically similar to a query, which is not always contextually relevant. It does not establish how fragments relate or what is missing, so the model must rebuild those relationships on every call. A context layer keeps the relationships between calls instead of rebuilding them.
The whitepaper argues no. Larger windows and more capable models help, but maintaining shared contextual understanding is a separate responsibility from assembling a window. That responsibility persists however large the window becomes, because understanding is built up over time rather than assembled in one call.
No. The layer is defined by its role, not its technology. It may use knowledge graphs, databases, event processing and inference models. What defines it is maintaining a coherent, evolving representation of relationships, grounded in a domain ontology and in human concerns and constraints, and making it available to applications.
What it means for judgment and for the people being cared for.
Clinicians and other responsible people do. Their concerns, objectives and constraints are authored by them and sit inside the model, so what counts as significant reflects what they care about. Human-authored boundaries continue to govern any response. The system can only propose.
They are kept as tentative. A context layer is a system of information, not a system of record: uncertainty, ambiguity and nuance are part of what it holds, not defects to resolve away. Patterns it surfaces can be raised to clinical validity by a clinician, and not before.
The whitepaper, The context window and the context layer, by Keith Deutsch, sets out the architecture in about 3,000 words. Shorter illustrations are in three articles that follow synthetic patients, and a talk by Keith explains the idea in under eleven minutes.
Last reviewed
