SDOH Data Quality

The degree to which all required SDOH data elements are present in the extracted record. Incomplete data leads to false negatives in risk stratification.

• Record-level: Is the patient's housing status documented at all?
• Attribute-level: If 'food insecurity' is flagged, is the severity or duration captured?
• Impact: A missing 'unemployed' status prevents a closed-loop referral for job assistance programs.

The extent to which extracted data conforms to defined value sets and syntactic rules. Invalid data breaks downstream FHIR interoperability.

• Format check: Is a Z59.0 (homelessness) code valid per the ICD-10-CM specification?
• Value domain: Does an extracted 'income level' fall within an acceptable range?
• Reference integrity: Does the linked FHIR SDOH Observation resource point to a valid patient ID?

The absence of logical contradictions between different data elements for the same patient across encounters. Inconsistent data erodes clinical trust.

• Cross-encounter: A note states 'patient is homeless' while a concurrent billing code lists a permanent address.
• Cross-system: The SDOH NLP pipeline extracts 'unemployed', but the EHR structured field shows an active employer.
• Resolution: Requires clinical validation rules engines to flag conflicts for human-in-the-loop review.

The latency between when an SDOH risk factor is documented in a clinical note and when it is available for action in the care management system.

• Real-time extraction: EHR-embedded NLP processes notes at the point of care to trigger a CDS Hook before the encounter ends.
• Batch processing: Nightly jobs may delay a critical referral by 24+ hours.
• Currency: A 'food insecure' status from 2018 is no longer timely for a 2024 care plan.

The measure of duplicate SDOH observations for a single patient. Redundant data inflates risk scores and wastes care manager resources.

• Deduplication logic: Identifying that 'homeless' in a progress note and 'Z59.0' in the problem list represent the same social risk episode.
• Entity resolution: Linking multiple mentions of the same community resource referral to a single closed-loop referral workflow.
• Impact: A single housing crisis should not generate five separate alerts.

The believability of extracted data within a clinical context, assessed through temporal and experiential logic.

• Temporality check: Is a 'history of homelessness' from 10 years ago being incorrectly flagged as a current active risk?
• Experiencer check: Is the 'alcohol abuse' mention attributed to the patient's spouse, not the patient, as confirmed by experiencer detection?
• Atemporal implausibility: A newborn with a documented 'unemployment' status triggers an immediate data quality review.

The foundational metric for data quality, measuring the correctness of an NLP system in identifying and classifying social determinant mentions. It is quantitatively assessed using precision (the fraction of extracted entities that are correct), recall (the fraction of all actual entities that were extracted), and F1-score (the harmonic mean of precision and recall). High accuracy is a prerequisite for valid downstream use in population health analytics.

Deterministic and probabilistic logic systems that act as a primary gatekeeper for data quality. These engines verify AI-extracted SDOH data against predefined clinical constraints, such as checking if an extracted 'food insecurity' mention is temporally consistent with a documented 'homelessness' status. They enforce completeness, validity, and consistency by flagging contradictory or implausible extractions for human review before the data enters a warehouse.

A critical contextual analysis technique that directly impacts the validity dimension of data quality. It distinguishes between affirmed ('patient is homeless'), negated ('patient denies homelessness'), and uncertain ('patient may be homeless') social risk findings. Failure to detect negation leads to false positives, corrupting datasets and inflating risk cohort counts. Modern approaches use contextual embeddings from models like Clinical BERT to resolve this scope.

The silent degradation of an extraction model's performance over time, representing a direct threat to long-term data quality. Drift occurs when the statistical properties of the input data change, often due to new screening tools, evolving clinical documentation patterns, or shifts in patient demographics. Continuous monitoring of F1-scores and data distributions is required to detect drift and trigger model retraining or recalibration.

A quality assurance workflow that is the final arbiter of data quality for high-stakes use cases. When an extraction model's confidence score falls below a defined threshold, the output is routed to a clinical reviewer for audit and correction. This process creates a validated, human-verified gold standard dataset that not only ensures immediate data quality but also provides the labeled data necessary for active learning and model improvement.

The foundational governance document for all SDOH data quality. These detailed instructions define the exact scope, entity types, and edge cases for human annotators labeling a gold standard corpus. High-quality guidelines ensure inter-annotator agreement (IAA), a measure of consistency between human labelers. Without rigorous, unambiguous guidelines, the training data itself is inconsistent, making high extraction accuracy impossible.