Source Attribution

Each AI-generated clinical statement or billing code is anchored to a precise character offset range (start and end position) in the source document. This granular linking enables reviewers to click a statement and instantly see the exact sentence or phrase that served as evidence.

• Maps structured outputs to unstructured text spans
• Supports multi-span attribution when evidence is distributed across paragraphs
• Enables one-click navigation from claim to source

Every attributed source link carries a calibrated confidence score that indicates the model's certainty about the evidence-to-output mapping. Low-confidence attributions are visually flagged in the review interface, allowing clinicians to prioritize verification of the most uncertain links.

• Color-coded confidence indicators (green/amber/red)
• Integrates with confidence threshold routing for task triage
• Reduces time spent verifying high-certainty extractions

Review interfaces present a side-by-side diff view where the AI-extracted statement appears alongside the attributed source text, with the relevant span highlighted. This visual alignment accelerates the human verification process by eliminating the need to manually search the original document.

• Highlights exact evidence spans in context
• Supports inline correction when attribution boundaries are incorrect
• Reduces cognitive load during high-volume review sessions

For clinical conclusions that synthesize information across multiple records—such as a prior authorization that references a progress note, a lab result, and a specialist consult—source attribution maintains a complete provenance chain. Each contributing document and its specific evidence spans are preserved in the audit trail.

• Tracks evidence across longitudinal patient records
• Supports FHIR Provenance resource mapping
• Enables full reconstruction of clinical reasoning paths

When a human reviewer adjusts an incorrect source attribution—such as expanding a span that missed critical context—the correction can be propagated to semantically similar extractions across the batch. This maintains attribution consistency without requiring manual correction of every identical error.

• Leverages correction propagation mechanisms
• Reduces review burden for systematic extraction errors
• Maintains consistency across large document sets

Every source attribution, including any human modifications, is recorded in an immutable audit trail that captures who verified the link, when it was reviewed, and what changes were made. This chain of custody is essential for compliance with payer audits and regulatory requirements.

• Timestamped reviewer identity and actions
• Supports discrepancy resolution workflows
• Provides defensible evidence for denied claims appeals

A visual comparison interface that highlights the specific textual, structural, or coding differences between an AI-generated output and a human-corrected version. In the context of source attribution, a diff view can show exactly which sentence in the original record was used to generate a claim, and how a reviewer's correction changes that linkage.

• Reduces time-to-verification by 40-60%
• Highlights character-level span changes
• Essential for audit trail completeness

A granular annotation task where a human reviewer adjusts the start and end character offsets of a highlighted medical entity in unstructured text. When source attribution links a diagnosis code to a specific paragraph, span correction allows the reviewer to tighten or expand the evidence window if the model's boundary detection was imprecise.

• Fixes extraction boundary errors
• Maintains precise provenance mapping
• Critical for FHIR resource accuracy

A chronological, tamper-proof record of all user interactions and system changes within a review interface. When combined with source attribution, the audit trail captures not just what was changed, but which evidence the reviewer consulted to make the correction, providing a complete chain of custody for clinical data modifications.

• Enables compliance verification
• Records evidence-to-decision pathway
• Supports retrospective quality audits

A predefined probability score below which a model's prediction is flagged for manual review. Source attribution directly supports threshold tuning by allowing reviewers to rapidly verify low-confidence extractions against the original text, enabling data-driven decisions about where to set the automation boundary.

• Balances automation rate vs. error risk
• Attribution links speed low-confidence review
• Enables dynamic threshold adjustment

A mechanism that automatically applies a single human correction to identical or semantically similar errors across a batch or downstream dataset. When a reviewer corrects a source attribution link for one instance of a recurring error pattern, propagation ensures all similar misattributions are fixed simultaneously.

• Maintains consistency across records
• Reduces repetitive manual effort
• Leverages attribution patterns for bulk fixes

A meticulously curated, high-quality set of ground truth clinical data used to benchmark model accuracy. Source attribution annotations in a golden dataset specify not just the correct extracted value, but the exact source span that justifies it, enabling precise evaluation of both extraction and evidence-linking performance.

• Evaluates extraction + attribution accuracy
• Used for reviewer norming sessions
• Establishes ground truth for model retraining