SDOH NLP Pipeline

Goes beyond simple keyword matching to identify social risk mentions within complex linguistic contexts. This component uses Clinical BERT embeddings to understand that 'homeless' refers to housing instability, not a general descriptor. It must differentiate between:

• Patient experiences: 'Patient is homeless'
• Family history: 'Father was homeless'
• Hypotheticals: 'If patient becomes homeless' This experiencer detection ensures only relevant patient-level risks are extracted.

A critical filtering layer that prevents false positives by accurately classifying assertion statuses. The system must distinguish:

• Affirmed: 'Patient reports food insecurity'
• Negated: 'Patient denies food insecurity'
• Uncertain: 'Possible housing instability' This module leverages syntactic dependency parsing and contextual cues to apply NegEx-style algorithms adapted for social determinants, ensuring only confirmed or probable risks populate structured data fields.

Classifies the chronological status of each extracted social risk to support accurate patient phenotyping. The pipeline assigns a temporal label:

• Current: 'Is currently unemployed'
• Historical: 'Was homeless 2 years ago'
• Future Risk: 'At risk of eviction' This temporality classification prevents outdated information from triggering inappropriate closed-loop referrals and ensures care managers act on active, modifiable risk factors.

Maps extracted natural language mentions to standardized terminologies for interoperability. The pipeline performs clinical entity linking to:

• ICD-10-CM Z-Codes: Z59.0 for homelessness, Z59.4 for food insecurity
• LOINC: Standardized screening panel codes
• SNOMED CT: Granular social circumstance concepts
• Gravity Project Value Sets: FHIR-compliant SDOH data elements This normalization enables SDOH interoperability across EHRs, HIEs, and community resource platforms.

Assigns a probabilistic confidence score to each extraction, enabling intelligent human-in-the-loop review workflows. High-confidence extractions (e.g., >0.95) flow directly into structured data. Low-confidence predictions are routed to a clinical reviewer interface for adjudication. This architecture supports active learning, where reviewer corrections are fed back to fine-tune the model, continuously improving SDOH extraction accuracy and mitigating model drift over time.

Augments patient-level extractions with area-level social risk context through geocoding. A patient's address is converted to precise geographic coordinates and linked to:

• Area Deprivation Index (ADI): Neighborhood socioeconomic disadvantage
• Social Vulnerability Index (SVI): Community resilience metrics This enrichment enables SDOH risk stratification that combines individual documented needs with population-level risk proxies, supporting health equity analytics and targeted resource allocation.

The foundational NLP subtask that identifies and categorizes specific mentions of social risk factors within clinical narratives.

Key extraction targets:

• Housing status: 'homeless', 'shelter', 'evicted'
• Food security: 'food insecure', 'skipping meals'
• Employment: 'unemployed', 'lost job', 'disability'
• Transportation: 'no car', 'missed appointment due to ride'

Technical approach: Transformer-based models fine-tuned on clinical corpora achieve F1 scores exceeding 0.85 for well-defined SDOH entity types. Span-level classification using BIO tagging schemes remains the dominant architecture.

A contextual analysis technique that distinguishes whether a social risk factor is affirmed, negated, or uncertain in clinical text.

Critical distinctions:

• Affirmed: 'Patient reports homelessness'
• Negated: 'Patient denies food insecurity'
• Historical: 'Patient was homeless 2 years ago'

Implementation: NegEx and ConText algorithms extended with SDOH-specific triggers. Modern approaches use contextual embeddings from Clinical BERT to capture long-range negation cues that rule-based systems miss.

A subset of diagnosis codes (Z55-Z65) used to document social determinants of health in a patient's structured medical record.

Code categories:

• Z55: Problems related to education and literacy
• Z56: Problems related to employment and unemployment
• Z57: Occupational exposure to risk factors
• Z59: Problems related to housing and economic circumstances
• Z60: Problems related to social environment
• Z65: Problems related to other psychosocial circumstances

Pipeline integration: Extracted SDOH entities are mapped to Z-codes for structured EHR insertion, enabling population health analytics and value-based care reporting.

A Fast Healthcare Interoperability Resources resource used to represent a specific, screened social risk finding in a standardized, exchangeable format.

Resource structure:

• code: The SDOH screening question or assessment
• value: The patient's response or score
• derivedFrom: Link to the source clinical note
• category: 'sdoh' category tag for discovery

Gravity Project alignment: Uses standardized LOINC codes for screening questions and SNOMED CT codes for answers, ensuring semantic interoperability across health systems and community-based organizations.

An NLP task that determines the chronological status of a social risk mention within a clinical narrative.

Temporal categories:

• Current: 'Patient is currently staying at a shelter'
• Historical: 'Patient experienced homelessness in 2019'
• Future risk: 'At risk of eviction next month'

Technical implementation: Combines date extraction with temporal relation classifiers. Transformer models with relative position encodings capture temporal context windows. Critical for distinguishing active needs from resolved social risks in population health dashboards.

A contextual NLP task that identifies who is experiencing the social risk mentioned in a clinical note.

Entity resolution targets:

• Patient: 'Patient reports food insecurity'
• Family member: 'Patient's spouse is unemployed'
• Caregiver: 'Daughter lost her job, impacting care'
• Household: 'Family is facing eviction'

Pipeline importance: Prevents false attribution of social risks to the index patient when the mention actually refers to a family member. Uses dependency parsing and coreference resolution to link risk mentions to the correct experiencer.