The intelligence substrate
A graph of care, not a database of records.
The Care Graph is the foundation every other part of careos sits on. It connects patients, episodes, tasks, medications, messages, and AI decisions into one longitudinal substrate — and it is the reason grounded, explainable automation is possible at all.
What the Care Graph connects
Nine entity classes. One substrate.
Every entity is typed, versioned, and linked. Every relationship is first-class. Every mutation produces an event.
Patients & proxies
Person identity, relationships, proxies, consent scope, contact preferences, and household links — structured, deduplicated, and versioned.
Episodes & care plans
Versioned plans, goals, interventions, reviews, and outcomes. Every plan is immutable once signed and every revision is addressable.
Tasks & escalations
Work items with SLA policies, escalation ladders, queue state, ownership history, and linkage back to the plan or signal that emitted them.
Medications
Catalogue (dm+d / RxNorm), reconciliation snapshots, administration, omission and refusal events with structured reason codes.
Messages
Secure threads scoped to patient, episode, appointment, or task. Channel state, delivery receipts, reply routing, proxy handling.
Observations
Vitals, symptoms, assessments, remote monitoring ingestion, threshold state, deterioration flags, and trend windows.
Documents
Forms, signatures, consent artefacts, clinical notes, and structured attachments with versioning and redaction support.
AI interactions
Every agent call with provenance: inputs, sources, model, prompt version, decision, reviewer, outcome, and downstream effects.
Events
Domain events for identity, referrals, scheduling, care execution, clinical signals, communications, compliance and AI — replayable, versioned, and streamable.
Why it matters
Records store the past. Graphs explain it.
Traditional systems store records. They capture what happened and put it on a screen. The Care Graph connects them — and connecting is the job. Most operational errors in care are not missing records, they are missing relationships between records that nobody stitched together in time.
AI reasons over structured relationships, not text chunks. When a careos agent answers a question, it walks the graph along typed edges to reach the nodes it needs. It does not guess. The same walk is loggable, auditable, and replayable — which is what makes automation safe in a clinical context.
The graph makes two-step visibility possible: this patient missed a dose AND has a high-risk observation AND has an open approval → flag. That class of detection is impossible on flat tables and painful in bolted-on BI. In the Care Graph it is a query.
This is the moat. Competitors can build workflows. They can build inboxes and dashboards. What they cannot copy quickly is the graph and the years of modelling that go into making it clinically correct.
What the graph enables
Features that are impossible without it.
Grounded AI
Agents reason over actual connected entities, not fuzzy retrieval over text chunks. Every AI output cites the graph nodes it depended on.
Root-cause analysis
Trace every outcome — good or bad — back through the sequence of events, tasks, and decisions that produced it. Pathway debuggability.
Novel features
Two-step risk detection, leakage prediction, cohort analysis, and operational heatmaps that are impossible on flat record stores.
Benchmark Network
Opt-in de-identified analytics across tenants — the first real comparability layer for ambulatory and community operators.
Cohort analysis
Slice active patient populations by pathway, risk, engagement and funding — then measure interventions against a matched baseline inside the graph.
Leakage prediction
Surface revenue and outcome leakage two steps ahead — missed contacts, stalled approvals, abandoned intakes and no-show sequences — from the connected event stream.
Technical characteristics
Built for regulated scale.
Longitudinal
The full patient timeline is queryable at scale with filters on episode, pathway, site, and time window.
Multi-tenant
Server-side isolation with attribute constraints. Tenant boundaries are enforced in the data plane, not application code.
Event-sourced
Every state change produces domain events. The current state is derived; the history is authoritative.
Auditable
Every mutation is linked to an Evidence Ledger entry with actor, source, policy reference, and reviewer if applicable.
Queryable
Supports graph traversal, time-series aggregation, and entity-level lookup through a single contract-first API.
Versioned schemas
Every entity, event and edge type is schema-versioned. Migrations are forward-compatible and downstream integrations are not silently broken.
The substrate
See the Care Graph in a walkthrough.
The best way to understand the graph is to see it query an actual pathway. We run walkthroughs for design partners and early-access teams.