SDOH Knowledge Graph

Grounds ambiguous social risk mentions to unique, standardized identifiers from authoritative ontologies.

• Maps 'homeless' to SNOMED CT 32911000 (Homeless) and ICD-10-CM Z59.0
• Links 'food insecurity' to Gravity Project value sets and LOINC screening codes
• Resolves 'unemployed' to UMLS C0041675 for cross-system interoperability
• Enables disambiguation between patient and family member experiences via experiencer detection

Leverages formal medical ontologies to derive implicit knowledge from explicit facts.

• Infers that Z59.0 (Homelessness) implies elevated risk for Z59.1 (Inadequate Housing)
• Propagates relationships through SNOMED CT hierarchical structures (is-a, has-finding)
• Applies temporal reasoning to distinguish current crises from historical resolved conditions
• Supports subsumption queries that retrieve all patients with any housing-related Z-code

Connects extracted patient needs directly to available intervention programs through a unified graph.

• Models Community-Based Organizations (CBOs) as nodes with service area, capacity, and eligibility properties
• Creates edges between SDOH findings and resource nodes based on semantic matching
• Enables closed-loop referral tracking by linking screening events to referral outcomes
• Supports geospatial queries by integrating Area Deprivation Index and Social Vulnerability Index data

Aggregates social risk data across time to construct a comprehensive patient trajectory.

• Connects discrete SDOH Observations into a temporal sequence for trend analysis
• Links social risk episodes to clinical outcomes, utilization patterns, and FHIR Encounter resources
• Maintains provenance edges to source documents, enabling full data lineage and auditability
• Supports risk stratification models by providing structured input features from the graph

Bridges disparate data silos by aligning social, clinical, and administrative concepts.

• Maps Gravity Project terminology to ICD-10-CM, SNOMED CT, and LOINC within a single graph
• Exposes queryable endpoints via FHIR SDOH Observation and FHIR Condition resources
• Integrates with USCDI SDOH Data Elements to ensure certified EHR compatibility
• Enables federated queries across institutional boundaries while preserving data governance policies

Provides transparent, auditable inference chains for every derived conclusion.

• Traces the exact path from a clinical note mention through entity linking to a resource recommendation
• Exposes negation detection results as graph properties, preventing false-positive referrals
• Records model confidence scores and human-in-the-loop corrections as versioned graph annotations
• Supports algorithmic fairness audits by making all inference steps visible and contestable

The fundamental data structure of a knowledge graph, expressing facts as subject-predicate-object statements. In an SDOH context, a triple might assert that 'Food Insecurity' (subject) 'is_associated_with' (predicate) 'ICD-10-CM Z59.4' (object). This structure enables machines to traverse relationships logically, moving from a patient's diagnosis code to a related social risk concept and then to a specific community intervention, forming the basis for automated reasoning.

A taxonomy is a hierarchical classification (e.g., SDOH > Economic Stability > Food Insecurity). An ontology is a richer, more formal representation that defines the nature of relationships between concepts. An SDOH ontology specifies not just that 'Homelessness' is a type of 'Housing Instability', but also its causal links to 'Emergency Department Utilization' and its required interventions from 'Community Resource Linkage' platforms, enabling complex inferencing.

The computational process of inferring new knowledge from existing graph data. Using forward-chaining or backward-chaining algorithms, a system can deduce a patient's unstated risks. For example, if a graph connects 'Unemployment' to 'Loss of Health Insurance' and that to 'Delayed Care', the engine can infer a high-risk profile for a patient with a documented Z56.0 code, even without an explicit 'Access to Care' problem on their record.

The critical process of disambiguating and linking textual mentions to unique nodes in the graph. The phrase 'homeless' in a clinical note must be resolved to the canonical node 'Homelessness (SNOMED CT: 32911000)' rather than a generic string. This ensures that all data—whether from structured Z-codes or unstructured NLP extraction—is unified into a single, queryable entity, preventing fragmentation and enabling accurate population-level analytics.

A technique that translates the discrete nodes and edges of a knowledge graph into a continuous, low-dimensional vector space. Algorithms like Node2Vec or GraphSAGE learn representations that preserve structural similarity. This allows for predictive tasks like link prediction—forecasting a patient's likely future social risks based on the vector proximity of their current profile to known risk trajectories in the embedding space.

The Resource Description Framework (RDF) serialization of FHIR data provides a native graph format for healthcare information. Representing an SDOH Observation as RDF allows it to be directly ingested into a knowledge graph without rigid relational mapping. This enables federated querying across disparate systems using SPARQL, where a single query can traverse from a patient's FHIR record to linked community resource directories and social vulnerability indices.