Named Entity Recognition for SDOH

Identifies the exact text boundaries of social risk mentions, which often span multiple words or phrases. Unlike simple keyword matching, the system must capture complex expressions like 'lost his job three months ago' or 'currently staying at a shelter' as single, coherent entities. This requires understanding syntactic dependencies and semantic roles to avoid fragmenting the mention into meaningless pieces.

Classifies extracted spans into a detailed SDOH ontology, moving beyond binary 'risk/no-risk' labels. The system must distinguish between:

• Housing Insecurity vs. Homelessness
• Food Insecurity vs. Nutritional Deficiency
• Unemployment vs. Underemployment
• Transportation Barriers vs. Financial Strain This granularity enables precise resource matching and population health analytics.

Determines the certainty and subject of each social risk mention. The system must accurately classify whether a risk is:

• Present: 'Patient reports homelessness'
• Absent: 'Patient denies food insecurity'
• Historical: 'Was unemployed last year'
• Family Member: 'Patient's spouse lost job' Misclassifying a negated or historical mention as a current risk leads to false positives and wasted intervention resources.

Links mentions of the same social risk across different clinical notes and encounters. For example, a 'housing issue' mentioned in a social work note must be resolved to the same entity as 'eviction notice' in a subsequent ED visit. This longitudinal linking builds a complete patient social history, preventing duplicate referrals and enabling trend analysis over time.

Detects social risks that are implied rather than explicitly stated. The system must recognize that 'patient is a veteran living alone on a fixed income' signals potential financial strain and social isolation, even without explicit keywords. This requires deep contextual embeddings and world knowledge to surface risks that clinicians document indirectly.

Maps extracted social risk entities to standard terminologies for interoperability and billing. The system must link mentions to:

• ICD-10-CM Z-Codes (Z59.0 for homelessness, Z59.4 for food insecurity)
• LOINC codes for screening instruments
• SNOMED CT concepts for social context
• Gravity Project value sets for FHIR exchange This mapping enables structured data capture and closed-loop referral workflows.

An automated sequence of natural language processing components designed to extract, classify, and structure social determinant risk factors from unstructured clinical narratives. A typical pipeline includes:

• Pre-processing: Sentence segmentation and tokenization of clinical notes
• Named Entity Recognition: Identifying spans of text mentioning social risks
• Negation Detection: Distinguishing affirmed risks from historical or negated ones
• Entity Linking: Mapping extracted mentions to standardized codes like ICD-10-CM Z-Codes (Z55-Z65)
• Structuring: Outputting FHIR-compliant observations for downstream use

A contextual analysis technique that distinguishes whether a social risk factor is present or absent in clinical text. This is critical because a note stating 'patient denies food insecurity' must not trigger a false positive alert. Modern approaches use:

• Rule-based systems: Regular expressions scoped to sentence boundaries
• Transformer-based classifiers: Fine-tuned models like Clinical BERT that learn linguistic cues of negation
• Scope resolution: Determining the exact span of text affected by a negation cue such as 'no evidence of' or 'denies'

Without robust negation detection, SDOH extraction systems produce noisy, untrustworthy data that undermines population health analytics.

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

• Z55: Problems related to education and literacy
• Z56: Problems related to employment and unemployment
• Z59: Problems related to housing and economic circumstances
• Z60: Problems related to social environment
• Z62: Problems related to upbringing

Despite their availability, Z-codes are severely underutilized in structured EHR data, making NLP extraction from free-text notes essential for capturing the true prevalence of social risk.

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

• Patient: 'Patient reports losing his job last month'
• Family member: 'Patient's spouse is currently unemployed'
• Caregiver: 'Daughter states she is facing eviction'

Misattributing a family member's housing instability to the patient corrupts the medical record and can lead to inappropriate referrals. This task often relies on dependency parsing and semantic role labeling to correctly assign the experiencer.

A Fast Healthcare Interoperability Resources resource used to represent a specific, screened social risk finding in a standardized, exchangeable format. The Gravity Project has defined standardized value sets for SDOH observations within FHIR, enabling:

• Interoperability: Seamless exchange of social risk data between EHRs, payers, and community-based organizations
• Coded representation: Mapping free-text mentions to LOINC and SNOMED CT codes
• Structured metadata: Capturing the screening tool used, the patient's response, and the temporal context

This standardization is foundational for closed-loop referral workflows and value-based care analytics.

An NLP task that determines the chronological status of a social risk mention. Accurate temporality prevents outdated information from driving clinical decisions:

• Current: 'Patient is currently staying at a shelter'
• Historical: 'Patient experienced homelessness 5 years ago'
• Future: 'Patient at risk of eviction next month'

Temporal reasoning often combines rule-based date extraction with transformer-based contextual classification to anchor risk mentions to specific timeframes, ensuring care teams act on active, not resolved, social needs.