Clinical Knowledge Base
How Patient Memory uses a curated vocabulary to infer relationships between conditions, medications, and labs that are not present in the source data.
What it is
Source records (FHIR bundles, CDA documents) tell you what a patient has. They rarely tell you how those facts relate. A FHIR bundle may contain both a Metformin prescription and a Type 2 Diabetes diagnosis without encoding any link between them; the relationship is implicit clinical knowledge that no EHR writes down.
The clinical knowledge base fills that gap. It is a curated set of vocabularies that the pipeline uses to infer typed relationships between entities after extraction and deduplication.
What it contains
The knowledge base is organized into four vocabularies:
Concepts
A canonical catalog of clinical concepts (conditions, medications, lab tests) with normalized display names, standard codes, and aliases. When the NLP normalization layer (Layer 2) needs to decide whether "T2DM" and "Type 2 Diabetes Mellitus" refer to the same concept, it looks here.
Drug indications
Maps medications to the conditions they are commonly prescribed for, with a strength level for each pairing. When Metformin and Type 2 Diabetes both appear in a patient's resolved graph, the pipeline emits a prescribed_for edge with a confidence score derived from the strength level.
| Strength | Confidence | Meaning |
|---|---|---|
strong | 0.90 | Primary indication. Drug is most often used for this condition |
moderate | 0.75 | Common secondary indication or off-label first-line use |
weak | 0.60 | Possible but not primary. Worth noting, lower confidence |
Condition relationships
Maps pairs of conditions to typed relationship edges. Covers complications, comorbidities, risk factors, manifestations, and conditions on the same disease axis.
| Relationship type | Example |
|---|---|
complication_of | Chronic kidney disease → Type 2 diabetes |
risk_factor_for | Hyperlipidemia → Coronary artery disease |
commonly_comorbid | Hypertension → Type 2 diabetes |
manifestation_of | Peripheral neuropathy → Type 2 diabetes |
same_disease_axis | Microalbuminuria → Chronic kidney disease |
Observation-conditions
Maps lab tests (by LOINC code) to the conditions they monitor. When HbA1c and Type 2 Diabetes both appear in the graph, the pipeline emits a monitors edge.
How it is used
The clinical knowledge graph is used during multiple steps of the Consolidate layer:
-
Concept normalization in entity resolution: The knowledge graph provides canonical concepts and aliases, allowing the app to unify different terms (e.g., "T2DM" and "Type 2 Diabetes Mellitus") into a single, normalized entity. This ensures consistency across extracted data from various sources.
-
Relationship inference after resolution: Once entities are resolved, the knowledge graph is used to infer clinical relationships (such as prescribed_for, complication_of, monitors) between them. This adds context and meaning to the patient graph, surfacing implicit links that are not present in the raw data.
-
Post-processing condition reclassification: The graph helps reclassify or group related conditions based on known relationships (e.g., recognizing that microalbuminuria is a manifestation of chronic kidney disease), improving the accuracy and clarity of the final patient summary.