Medisight Patient Profiles

Graph layer

How Graphify contributes graph structure

Graphify turns mixed patient files into an explainable knowledge base: evidence nodes at the top, clinical state nodes in the middle, and durable patient memory at the bottom.

What gets built

Nodes for concepts, events, facts, episodes, documents, and source spans.
Edges for support, contradiction, sequence, provenance, and encounter relationships.
Graph state that lets the system traverse from a patient concept to the evidence trail behind it.

Where graph structure helps

Follow transitions like “what changed after atrial fibrillation was first confirmed?”
Inspect disagreement chains such as denial versus later confirmation or tolerance events.
Group temporally related activity into episodes rather than treating the chart like a flat bag of files.

What graph structure does not solve by itself

Semantic recall for loosely phrased patient-history questions.
Ranking large numbers of candidate evidence spans by contextual similarity.
The final patient-profile computation without a contradiction and temporal resolution layer.

Vector layer

How Qdrant contributes vector intelligence

The vector layer solves the "too much chart text" problem. It searches across huge numbers of spans, summaries, and notes, then narrows them to a small set of semantically relevant evidence.

Qdrant pipeline

Chunk or span extraction from clinical files.
Embedding generation for raw evidence, episode summaries, and selected fact layers.
Qdrant nearest-neighbor search to recover semantically related evidence.
Candidate set returned to the hybrid retrieval controller.

What vector intelligence adds

Semantic recall across inconsistent chart phrasing and long-horizon patient history.
Evidence retrieval that is not limited to exact symbols or pre-linked graph paths.
A broad candidate set that can be narrowed later by temporal, provenance, and contradiction-aware ranking.

Important: the patent posture should not rise or fall on vector search. The vector database is a subcomponent. The defensible core is still contradiction-resolved patient-state compilation and retrieval control.

Fusion logic

How both systems work together

The fusion flow is graph first, vector second. The graph layer finds the right patient state and time window, then the vector layer drills into raw text to fetch the most specific evidence.

Graph inputs to the fusion layer

Fact history, contradiction edges, and chronology.
Episode boundaries and related event clusters.
Profile-slot candidates such as active conditions, current medications, and allergies.

Vector inputs to the fusion layer

Semantically similar raw evidence spans from Qdrant.
Recovered text that may not yet be represented in graph traversal results.
Supporting context for source-of-truth, ever-had, transition, and denial-oriented questions.

Fusion outputs

Contradiction-resolved facts with support and contradiction linkage.
Intent-conditioned retrieval plans driven by patient state.
Coverage-aware evidence sets ready for citation-grounded answers.

Hybrid ranking signals

vector similarity slot match score temporal relevance episode relevance source authority corroboration bonus contradiction penalty extraction confidence concept persistence

Patient profile output

What precise patient profiles actually contain

The end product is not a generic chart summary. It is a patient-state representation that can answer history questions with traceable evidence and explicit state logic.

Profile objects

PatientEvent objects extracted from evidence spans.
PatientFact objects with current, historical, resolved, ruled-out, or uncertain state.
Episode memory for admissions, procedures, treatment cycles, and related change windows.
Profile slots for active conditions, medications, allergies, procedures, imaging, and pathology.

Question types the system should handle

Current state: what is active now?
Ever had: has this concept ever been supported?
Transition after event: what changed after the first confirmed event?
Rule out or denial: is the concept denied, contradicted, or copy-forward noise?

Why profiles become more precise

Graph traversal preserves chronology and contradiction paths.
Qdrant preserves semantic recall over raw chart evidence.
The patient-state compiler reconciles both into explicit facts and evidence-backed outputs.

Technical posture for the patent story

The strongest story is still a contradiction-resolved patient-state compiler with intent-conditioned hybrid retrieval. Graphify and Qdrant should be shown as enabling subsystems that feed the compiler, not as the invention by themselves.

End-to-end architecture

Build patient memory once, then answer with graph-first hybrid retrieval

The final system has two phases: build the patient memory from mixed repositories, then answer queries by traversing graph structure first and using vector search for exact evidence retrieval.

Graph-first planning

Map the question to a concept, slot, or episode.

Traverse chronology and contradiction structure.

Constrain the time window and evidence zone before text search.

Produce a targeted retrieval plan instead of searching blindly.

Patient-state resolution

Temporal normalization: event_time vs recorded_time

Targeted vector evidence retrieval under graph constraints

Contradiction resolver and fact-state computation

Episode builder and profile-slot compiler

Coverage-aware answer controller with citations

Vector drill-down

Search across very large text collections without exact-match dependence.

Use graph-derived filters to narrow patient, concept, and time scope.

Return the best supporting note spans, summaries, and fact evidence.

Hand a small high-value evidence set to the ranking controller.

Output Precise patient profiles

Facts, episodes, slots, citations, and confidence linked back to source evidence.

Why this matters More than graph-plus-vector search

The patient profile is computed state, not just retrieved text.