Span Correction

The core mechanism involves modifying start and end character offsets of an entity span within unstructured text. Unlike simple label correction, this task requires pixel-perfect precision to ensure the extracted string exactly matches the intended clinical concept.

• Example: Changing a span from 'metformin 500' to 'metformin 500 mg' to capture the complete medication dosage
• Mechanism: Reviewers click and drag or use keyboard shortcuts to expand or contract the highlighted region
• Criticality: A single character offset error can change 'family history of cancer' to 'history of cancer', altering clinical meaning

Span errors fall into distinct taxonomical categories that drive targeted model retraining and reviewer training interventions.

• Under-extraction: The span captures only a fragment of the entity, e.g., 'carcinoma' instead of 'squamous cell carcinoma'
• Over-extraction: The span includes extraneous tokens, e.g., 'the patient has diabetes' instead of 'diabetes'
• Offset shift: The span is correctly sized but anchored at the wrong position, capturing adjacent but irrelevant text
• Boundary ambiguity: Occurs with nested or overlapping entities, such as 'left lower lobe pneumonia' where 'left lower lobe' is an anatomical site and 'pneumonia' is a finding

Effective span correction demands specialized UI components that minimize cognitive load and maximize throughput.

• Token-level highlighting: Text is rendered with visible token boundaries so reviewers can snap adjustments to word edges
• Diff visualization: The original model-predicted span is shown alongside the reviewer's corrected span with strikethrough and highlight formatting
• Keyboard-first navigation: Shortcuts like Alt+Left/Right to expand or contract spans by one token enable rapid correction without mouse dependency
• Contextual preview: A side panel displays the full sentence or paragraph to prevent reviewers from losing document context during granular edits

Span correction directly determines the quality of extracted structured fields that populate EHR systems, FHIR resources, and analytics databases.

• Medication extraction: An incorrect span for 'lisinopril 10 mg daily' could result in a missing dose or frequency in the structured medication list
• Problem list population: Boundary errors on diagnosis spans cause incomplete or inaccurate SNOMED CT code mappings
• Temporal reasoning: Span boundaries that exclude temporal modifiers like 'history of' or 'new onset' corrupt clinical timelines
• Cohort identification: Span errors in inclusion criteria extraction lead to incorrect patient cohort assembly for clinical trials

Span-level annotation requires rigorous QA protocols because traditional label-level agreement metrics are insufficient for boundary precision.

• Exact match IAA: Requires identical start and end offsets between annotators, a stricter standard than label agreement
• Partial overlap scoring: Metrics like Jaccard index or Dice coefficient quantify the degree of span overlap when exact matches are rare
• Adjudication triggers: Cases where annotator spans diverge beyond a predefined character threshold are automatically escalated to a senior reviewer
• Golden span datasets: Curated reference sets with pixel-perfect boundaries serve as calibration benchmarks for both models and human reviewers

A single span correction can be programmatically propagated to identical or semantically similar errors across a document or batch, dramatically reducing review burden.

• Exact match propagation: The same string occurring elsewhere in the document is automatically corrected with the same boundary adjustment
• Pattern-based propagation: Regular expression or embedding similarity triggers apply the correction to structurally analogous spans
• Reviewer confirmation gate: Propagated corrections are surfaced in a lightweight confirmation UI rather than applied silently, maintaining human oversight
• Audit trail integrity: Each propagated correction is logged with its origin span and propagation rule for full traceability

A visual comparison interface that highlights the specific textual, structural, or coding differences between an AI-generated output and a human-corrected version to accelerate validation. In the context of span correction, a diff view visually contrasts the original model-predicted character offsets with the reviewer's adjusted boundaries, using strikethroughs and highlights to make the modification immediately apparent without requiring the reviewer to mentally map index values.

A statistical measure that quantifies the degree of consensus among multiple human reviewers, used to establish ground truth reliability. For span-level annotation, standard token-level metrics like Cohen's Kappa are insufficient. Instead, specialized overlap measures such as F1-score on exact boundary matching or intersection-over-union (IoU) thresholds are employed to determine if two annotators agree on the precise character offsets of a clinical entity.

A mechanism that automatically applies a single human correction to identical or semantically similar errors across a batch or downstream dataset to maintain consistency. When a reviewer fixes the span of a specific medication mention, the system can propagate that boundary adjustment to all other occurrences of the same string in the document, dramatically reducing the review burden and ensuring uniform data quality without requiring redundant manual corrections.

A feature that directly links an AI-generated clinical statement or code to the exact sentence or paragraph in the original medical record, enabling rapid evidence verification. During span correction, source attribution provides a one-click navigation mechanism that scrolls the source document to the precise location of the highlighted entity, allowing the reviewer to visually confirm whether the model's character offsets correctly capture the full clinical concept or require boundary adjustment.

A structured classification system of potential model failure modes used by reviewers to tag corrections, enabling granular performance analysis and targeted model retraining. For span correction tasks, common error categories include:

• Left Boundary Error: The start offset is incorrect, often truncating a modifier
• Right Boundary Error: The end offset is incorrect, often including trailing punctuation
• Over-extraction: The span captures text beyond the target entity
• Under-extraction: The span captures only a fragment of the target entity

A meticulously curated, high-quality set of ground truth clinical data used as a benchmark to evaluate model accuracy and calibrate reviewer proficiency. For span-level tasks, a golden dataset contains exact character-level annotations validated through a rigorous adjudication workflow where multiple expert annotators achieve high IAA. This dataset serves as the definitive reference for measuring whether a model's boundary predictions or a reviewer's corrections meet the required precision standard.