SDOH Risk Stratification

Combines individual-level extracted SDOH data (e.g., housing instability from clinical notes) with area-level indices like the Area Deprivation Index (ADI) and Social Vulnerability Index (SVI).

• Creates a composite risk score that captures both personal circumstance and environmental context
• Prevents ecological fallacy by not relying solely on zip-code-level proxies
• Enables granular segmentation beyond simple binary flags

Employs gradient-boosted trees (XGBoost, LightGBM) and logistic regression models trained on structured Z-codes, FHIR SDOH Observations, and NLP-extracted risk mentions.

• Models predict outcomes like 30-day readmission, ED utilization, or care gap persistence
• Feature importance analysis reveals which social factors most influence utilization
• Models are validated across demographic subgroups to assess algorithmic fairness

Incorporates temporality classification from NLP pipelines to distinguish between current, historical, and resolved social risks.

• A past episode of homelessness carries different weight than an active housing crisis
• Enables time-sensitive stratification that reflects changing patient circumstances
• Supports longitudinal tracking of risk trajectory over multiple encounters

Produces actionable risk segments typically organized into 3-5 tiers (e.g., Low, Moderate, High, Critical) mapped to intervention intensity.

• Critical tier: Immediate case management and closed-loop referral activation
• High tier: Proactive outreach and community resource linkage
• Moderate tier: Automated resource information delivery
• Each tier triggers specific workflows within the EHR via CDS Hooks integration

Risk stratification outputs directly feed into closed-loop referral systems that match patients to community-based organizations.

• High-risk patients are automatically queued for care coordinator review
• Referral outcomes are tracked and fed back into the model as training labels
• Creates a continuous learning loop that refines stratification accuracy over time

Continuous monitoring for SDOH model drift caused by changing documentation patterns, new screening tools, or population shifts.

• Statistical process control charts track feature distributions and prediction stability
• Automated retraining triggers when performance degrades below threshold
• Human-in-the-loop review audits low-confidence predictions to maintain quality

The practice of evaluating and mitigating bias in predictive models that segment patients by social risk. Key considerations include:

• Representation bias: Are training data reflective of the target population?
• Measurement bias: Are proxies like SVI accurate across all groups?
• Allocative harm: Does the model unfairly deny resources to certain subgroups? Fairness audits ensure risk stratification does not perpetuate or amplify existing health disparities.

An automated workflow that tracks a patient's journey from a positive social risk screening through to a confirmed connection with a community-based service provider. The loop includes:

1. Identify: Risk stratification flags the patient.
2. Refer: An electronic referral is sent to a community organization.
3. Confirm: The organization acknowledges receipt.
4. Close: The organization reports back on service delivery. This ensures accountability and measures the effectiveness of interventions.