SDOH Phenotyping

The defining characteristic of phenotyping is the synthesis of heterogeneous data types. Algorithms simultaneously process structured ICD-10 Z-codes (Z59 for housing, Z63 for social support) and unstructured clinical narratives. This fusion compensates for the well-documented under-coding of Z-codes, where a patient's housing instability is discussed extensively in a social work note but never formally diagnosed. The phenotype is derived from the union of these signals.

Phenotyping algorithms must classify the chronological status of a social risk. A mention of 'evicted last year' requires a different clinical response than 'eviction notice received today.' Advanced systems use temporality classification to distinguish between:

• Historical: Resolved past events
• Active: Current, ongoing stressors
• Future: Impending risks This temporal dimension is critical for prioritizing interventions and avoiding alert fatigue.

A critical technical challenge is determining who is experiencing the social risk. The statement 'patient's husband lost his job' indicates household financial strain but not the patient's direct unemployment. Phenotyping systems use dependency parsing and entity linking to attribute the risk to the correct individual—patient, family member, or caregiver—ensuring the patient's phenotype is not incorrectly labeled with a relative's social determinant.

Accurate phenotyping requires robust negation detection to distinguish affirmed risks from ruled-out ones. The phrase 'patient denies food insecurity' must negate the phenotype, while 'patient reports running out of food' must affirm it. More complex is uncertainty detection: 'patient is worried about making rent' suggests a risk that has not yet materialized. These linguistic nuances are processed using contextual embeddings from models like Clinical BERT.

Patient-level data is often enriched with area-level indices to create a composite phenotype. After geocoding a patient's address, the algorithm appends metrics like the Area Deprivation Index (ADI) or Social Vulnerability Index (SVI). This allows a system to infer risk even when no direct mention exists in the chart—a patient living in a neighborhood with a high SVI score for housing-transportation vulnerability receives a contextual risk flag.

A phenotype is not a static label but a dynamic trajectory. Algorithms track social risk mentions over multiple encounters to identify patterns:

• Transient: A single, resolved episode of need
• Recurrent: Cyclical crises (e.g., seasonal utility shutoffs)
• Chronic: Persistent, unresolved structural barriers This longitudinal analysis distinguishes a patient with a single past housing crisis from one with chronic housing instability, driving different care management pathways.

An automated sequence of natural language processing components that transforms unstructured clinical narratives into structured social risk profiles. A typical pipeline includes:

• Sentence segmentation and tokenization
• Named Entity Recognition for social risk factors
• Negation detection to exclude historical or denied risks
• Experiencer detection to confirm the patient is the subject
• Temporality classification to determine if the need is current

This pipeline is the core engine that feeds the phenotyping logic.

The application of predictive models to segment a patient population by their level of social risk. Phenotyping outputs serve as input features for these models, which assign each patient a risk tier (e.g., low, medium, high).

• Enables targeted intervention allocation
• Supports value-based care contract performance
• Integrates clinical and social risk for holistic scoring

Effective stratification depends on the completeness and accuracy of the underlying phenotype definitions.

The systematic evaluation and mitigation of bias in models that extract or predict social risk. Phenotyping algorithms must be audited to ensure they do not under-identify needs in certain demographic groups due to documentation disparities.

• Disparate impact testing across race, ethnicity, and language
• Equal opportunity metrics for true positive rates
• Annotation guideline audits for cultural bias

Fairness is a critical governance requirement for any production SDOH phenotyping system.

An automated workflow that tracks a patient's journey from a positive social risk phenotype through to a confirmed connection with a community-based service provider. The phenotyping output triggers the referral, and the loop is closed when the community-based organization confirms service delivery.

• Integrates with community resource linkage platforms
• Tracks referral status: sent, accepted, fulfilled
• Provides audit trail for health equity reporting

This workflow transforms phenotype identification into measurable social care outcomes.