Algorithmic Fairness for SDOH

Quantitative evaluation of whether an SDOH extraction model performs differently across protected demographic subgroups. This involves calculating performance metrics (precision, recall, F1-score) stratified by race, ethnicity, gender, and socioeconomic status.

• Key Metric: Equalized Odds Difference — measures whether false positive and false negative rates are balanced across groups
• Threshold: A difference exceeding 5% between subgroups typically triggers a fairness audit
• Example: A model achieving 92% recall for housing instability mentions in white patients but only 78% for Black patients exhibits actionable disparate impact

Systematic underrepresentation or overrepresentation of specific populations in the annotated corpora used to train SDOH extraction models. This bias originates from historical healthcare access disparities and documentation patterns.

• Source: Patients with fewer clinical encounters generate less training data, creating a feedback loop of invisibility
• Detection: Compare demographic distributions in training data against census or health system population benchmarks
• Mitigation: Oversampling underrepresented groups, synthetic data generation for rare SDOH mentions, and stratified sampling during corpus construction

Bias introduced during the human annotation process when labelers apply subjective judgment to ambiguous SDOH mentions. Annotation guidelines that lack cultural context can encode annotator assumptions into ground truth.

• Manifestation: Inconsistent labeling of housing instability when annotators interpret 'couch surfing' differently across cultural contexts
• Mitigation: Diverse annotation teams, bias-awareness training, and inter-annotator agreement analysis disaggregated by document demographics
• Measurement: Fleiss' Kappa calculated separately for documents mentioning different racial or ethnic groups to detect systematic labeling discrepancies

The risk that SDOH extraction models rely on proxy variables that correlate with protected attributes rather than the actual social risk factor. ZIP code, for example, can serve as a proxy for race due to residential segregation.

• Problem: A model using geocoded address features may learn racial associations instead of true housing quality indicators
• Detection: Feature importance analysis and ablation studies to identify which input features drive predictions for different subgroups
• Solution: Adversarial debiasing techniques that penalize the model for encoding protected attribute information in its internal representations

Evaluation of model performance across intersecting identity dimensions rather than single-axis demographic categories. A model may appear fair for 'women' and 'Black patients' separately while failing for Black women specifically.

• Methodology: Subgroup analysis across all combinations of race, gender, age, and primary language
• Example Finding: An SDOH model showing 90% accuracy overall but only 65% for elderly non-English-speaking patients
• Tooling: Fairness-aware evaluation frameworks that compute metrics for all intersectional subgroups with sufficient sample size

The downstream consequence when biased SDOH extraction leads to differential resource allocation. If a model systematically under-identifies food insecurity in certain populations, those patients receive fewer nutritional intervention referrals.

• Mechanism: Extraction errors propagate into risk stratification scores, which determine eligibility for care management programs
• Monitoring: Track closed-loop referral completion rates stratified by demographic groups to detect allocative disparities
• Governance: Human-in-the-loop review with fairness-aware sampling that prioritizes audit of predictions for historically marginalized groups

A quantitative evaluation methodology that measures whether an SDOH extraction model's performance differs significantly across protected demographic subgroups. The 80% rule is a common benchmark: the model's F1-score for a disadvantaged group should be at least 80% of the F1-score for the advantaged group.

• Compares precision and recall parity across race, ethnicity, and language groups
• Uses bootstrap confidence intervals to determine statistical significance
• Applied at the entity level: housing instability extraction may be accurate for English speakers but degrade for Spanish-dominant patients
• Tools: Fairlearn, Aequitas, and custom scikit-learn fairness metrics

A training-time intervention that generates synthetic clinical notes where only the demographic attribute is altered while the clinical and social risk content remains identical. This teaches the model to ignore spurious correlations between identity and outcome.

• Example: '67yo Black male presents with housing instability' → '67yo White male presents with housing instability'
• Preserves clinical fidelity while breaking demographic shortcuts
• Reduces stereotype amplification in transformer-based SDOH extractors
• Implemented using controlled text generation or template-based substitution

An in-processing technique where a gradient reversal layer is inserted into the SDOH extraction model architecture. A secondary adversary network attempts to predict the protected attribute from the model's internal representations, while the primary model is trained to simultaneously maximize extraction accuracy and minimize the adversary's success.

• Forces the encoder to learn demographic-invariant features
• Particularly effective for Clinical BERT fine-tuning pipelines
• Requires careful tuning of the adversarial weight hyperparameter to balance fairness and accuracy
• Can be applied during domain adaptation to new hospital systems

A post-processing calibration strategy where the confidence threshold for flagging an SDOH mention as positive is tuned independently for each demographic subgroup to achieve equalized error rates. Rather than a single global threshold, the system uses subgroup-specific operating points.

• Optimizes for equal opportunity or equalized odds constraints
• Does not require retraining the underlying model
• Addresses calibration drift where a model is systematically overconfident or underconfident for certain populations
• Implemented using ROC curve analysis per subgroup on a held-out validation set

A data governance intervention that ensures the human annotators labeling SDOH training data represent diverse lived experiences, linguistic backgrounds, and cultural competencies. Homogeneous annotator pools introduce implicit bias into gold-standard labels.

• Recruit annotators with community health expertise and multilingual fluency
• Use adjudication protocols where disagreements are resolved by a diverse panel, not a single authority
• Measure inter-annotator agreement by subgroup to detect systematic labeling disparities
• Critical for nuanced SDOH concepts like 'unstable housing' which have cultural variability in documentation

An MLOps practice that embeds automated fairness metric computation into the SDOH model's production pipeline. Dashboards track drift in demographic parity and equal opportunity over time, triggering alerts when disparities exceed predefined thresholds.

• Integrates with model observability platforms like Evidently AI or custom Prometheus metrics
• Monitors for emergent bias as documentation patterns shift with new screening initiatives
• Includes slice-based evaluation on intersecting subgroups (e.g., race × language × payer type)
• Feeds into a human-in-the-loop review queue that prioritizes audits for high-disparity slices

The degradation of an SDOH extraction model's performance over time due to changes in clinical documentation patterns, screening tools, or patient population characteristics. Model drift can silently introduce or amplify bias as the relationship between language and social risk shifts. Continuous monitoring against fairness metrics is essential to detect when a model's predictions become less accurate for specific subgroups, triggering retraining or recalibration.

The framework of policies, roles, and processes that ensures the ethical, secure, and high-quality collection, management, and use of sensitive social determinant data. A robust data governance program is the foundational prerequisite for algorithmic fairness, establishing clear accountability for bias testing, defining acceptable use, and mandating transparent documentation of model limitations and intended use.

A detailed instruction manual for human annotators that defines the scope, entity types, and edge cases for labeling social determinants of health in a gold standard corpus. Annotation guidelines are a critical source of upstream bias. If guidelines encode subjective interpretations of social risk or fail to address dialectal variations, the resulting training data will bake those biases into the model.

A composite metric using census data to measure a community's resilience to external stressors, often used as a proxy for patient-level social risk. Using area-level indices like the SVI as a feature requires careful fairness analysis. These proxies can introduce ecological fallacy, where group-level characteristics are incorrectly attributed to an individual, potentially stereotyping patients based solely on their zip code.

A contextual NLP task that identifies who is experiencing the social risk mentioned in a clinical note, distinguishing between the patient and a family member or caregiver. Experiencer detection is a fairness-critical subtask. Misattributing a family member's housing instability to the patient creates a false positive that can lead to inappropriate interventions and a biased overestimation of risk for certain demographics.

A quality assurance workflow where a clinical reviewer audits and corrects AI-extracted SDOH data, typically focusing on low-confidence predictions flagged by the model. This process is a primary mechanism for procedural fairness, allowing for human override of algorithmic decisions. The design of the review interface and the selection of which predictions to flag directly impact the fairness of the final output.