Discrepancy Resolution

A visual comparison interface that highlights specific textual, structural, or coding differences between an AI-generated output and a human-corrected version. This component accelerates validation by:

• Displaying side-by-side source and extraction with color-coded deltas
• Highlighting character-level span differences for boundary errors
• Enabling inline editing directly within the diff visualization
• Supporting keyboard shortcuts for rapid accept/reject decisions

Effective diff views reduce time-to-correction by 40-60% compared to manual cross-referencing.

A structured escalation process where a third, often more senior reviewer resolves a discrepancy between two initial annotators to establish a final reference standard. Key characteristics include:

• Blinded review to prevent anchoring bias on prior decisions
• Tie-breaking logic that routes unresolved conflicts to domain experts
• Rationale capture requiring adjudicators to document their reasoning
• Feedback loops that propagate adjudication decisions to improve initial reviewer calibration

This workflow is essential for establishing ground truth reliability in ambiguous clinical cases where inter-annotator agreement falls below acceptable thresholds.

A feature that directly links an AI-generated clinical statement or code to the exact sentence or paragraph in the original medical record. This component enables:

• One-click evidence verification by highlighting the source span
• Confidence calibration by exposing the textual basis for each extraction
• Rapid discrepancy identification when extracted data contradicts source text
• Audit trail completeness by preserving provenance metadata

Source attribution transforms the review process from a trust-based model to an evidence-based verification workflow, reducing cognitive load and verification time.

A mechanism that automatically applies a single human correction to identical or semantically similar errors across a batch or downstream dataset. This component maintains consistency by:

• Detecting duplicate extraction patterns using fuzzy matching
• Applying corrections to all affected records in the review queue
• Flagging propagated corrections for optional secondary verification
• Updating downstream FHIR resources and coded data elements

Correction propagation prevents the inefficiency of manually fixing the same systematic error across hundreds of documents, directly improving straight-through processing rates.

A structured classification system of potential model failure modes used by reviewers to tag corrections. This enables granular performance analysis by categorizing discrepancies into:

• Span boundary errors: Incorrect start/end offsets for entity extraction
• Entity misclassification: Wrong clinical concept type assigned
• Negation failures: Missed or incorrectly applied negation cues
• Relationship errors: Incorrect linking between related entities
• Hallucination: Extracted data with no source document basis

Tagged error distributions feed directly into targeted model retraining and reviewer calibration sessions, creating a continuous improvement loop.

A specialized interface component designed to visually align and compare two conflicting data sets for manual merging. Common applications include:

• Aligning AI-derived medication lists with existing EHR records
• Comparing problem list extractions against coded diagnoses
• Reconciling allergy documentation across multiple source documents

The reconciliation UI presents conflicting values side-by-side with source attribution, allowing reviewers to select the correct value, merge information, or flag for clinical review. This component is critical for medication reconciliation automation and problem list maintenance.