Correction Propagation

The engine uses dense vector embeddings to identify not just exact string matches, but semantically identical errors. If a reviewer corrects 'HTN' to 'Hypertension' in one note, the system finds all contextual variants like 'htn' or 'high blood pressure' that map to the same concept.

• Uses cosine similarity thresholds on clinical embeddings
• Prevents over-propagation by respecting negation contexts
• Distinguishes between identical surface forms with different meanings

When processing a bulk upload of clinical documents, a single correction is applied transactionally across the entire batch to ensure referential integrity. This prevents a scenario where a patient's allergy is updated in one document but remains erroneous in another.

• Operates within a logical batch boundary
• Maintains an immutable audit trail of cascaded changes
• Supports rollback of propagated corrections if a mistake is identified

Propagation extends beyond the immediate text to dependent structured fields. Correcting a medication name automatically updates the linked RxNorm code, dosage frequency, and route of administration in the target FHIR resource.

• Triggers re-evaluation of conditional logic rules
• Updates derived quality measures and clinical decision support triggers
• Prevents orphaned references in relational data models

Corrections are not applied blindly. The system only propagates a fix to instances where the original model confidence fell below a configurable threshold. High-confidence predictions that differ from the human correction are flagged for a separate adjudication workflow rather than being overwritten.

• Prevents silent overwrites of potentially correct model outputs
• Uses calibrated probability scores to gate propagation
• Generates a discrepancy report for conflicting high-confidence items

The system classifies the type of correction made to determine the propagation strategy. A span boundary adjustment propagates differently than a concept normalization or a negation flip.

• Boundary fixes adjust character offsets in similar contexts
• Concept normalizations map to standardized ontology IDs
• Negation flips trigger re-evaluation of all related clinical assertions

Before a correction is committed across the dataset, the reviewer is shown a diff view of all affected instances. This allows a human to deselect specific items where propagation would be contextually inappropriate, maintaining ultimate human authority.

• Displays a side-by-side comparison of original vs. propagated text
• Highlights outlier instances where semantic context differs
• Requires explicit confirmation before batch mutation