Source Attribution

The most granular form of attribution, linking an extracted data point to the exact sentence and document offset from which it was derived. This allows a clinical reviewer to click a medication entry and immediately see the highlighted source text, reducing verification time from minutes to seconds. Span annotation techniques capture character-level positions, while document object identifiers preserve the source file path and page number for PDF-based records.

When a single medication entry is corroborated by multiple sources—such as a Best Possible Medication History (BPMH) interview, a pharmacy dispense record, and a specialist's note—each contributing source is individually attributed. The system maintains a confidence weighting per source, allowing reviewers to see not just what was extracted, but from where and with what degree of corroboration. This is critical for resolving unintentional discrepancies.

Every extraction event is logged with a cryptographically verifiable record of its origin, timestamp, and the model version that produced it. This data provenance chain ensures that if a medication list is later questioned during a root cause analysis of an adverse drug event, investigators can trace exactly which sentence in which document generated each entry. This satisfies the evidentiary requirements of patient safety organizations.

Source attribution is tightly coupled with confidence thresholding. When a model extracts a medication with low certainty, the attribution link is flagged for mandatory human-in-the-loop (HITL) review. The reviewer sees not only the extracted value but the exact source context, enabling rapid adjudication. High-confidence extractions with clear, unambiguous sources may bypass manual review entirely, optimizing pharmacist workflow.

When the same medication appears across multiple documents—a discharge summary, an admission note, and a pharmacy fill record—source attribution links each mention to a unified RxNorm concept while preserving the distinct origin of each instance. This enables temporal reasoning across a patient's Medication History Longitudinal Record, showing when and where each medication was introduced, modified, or discontinued.

Source attribution preserves the surrounding linguistic context, which is essential for interpreting negation and uncertainty. An extraction of 'lisinopril' is linked to its source sentence, allowing the reviewer to see whether the text stated 'continue lisinopril,' 'discontinued lisinopril due to cough,' or 'no lisinopril.' This prevents omission errors where a negated medication is incorrectly added to the active list, a common failure mode of naive extraction systems.

The documented audit trail that tracks the origin, source system, and transformation history of a specific medication data point. Unlike simple source attribution which links to a sentence, data provenance captures the full lineage—from the original EHR database field through any ETL transformations to the final structured output. This ensures traceability required for clinical safety validation and regulatory compliance.

A probabilistic gate that routes AI-extracted medication data for human review only when the model's prediction score falls below a predefined certainty level. This mechanism directly complements source attribution by ensuring that low-confidence extractions—where the source text is ambiguous or contradictory—are flagged with their originating text for pharmacist verification. It optimizes the balance between automation and safety.

A system design paradigm where clinical pharmacists or technicians review, correct, and approve the output of an AI medication reconciliation engine before it is finalized. Source attribution is the critical interface for HITL—it provides the direct link from each extracted medication to the original text, enabling rapid verification without requiring the reviewer to re-read the entire chart.

Deterministic constraints and post-processing rules applied to LLM outputs to prevent the generation of plausible-sounding but factually non-existent medication names or dosages. Source attribution serves as a primary guardrail: if an extracted entity cannot be linked back to a specific source span, it is automatically flagged as a potential hallucination and suppressed from the final output.

A regular expression-based NLP algorithm specifically designed to identify whether a clinical finding mentioned in narrative text has been negated by the clinician. When combined with source attribution, NegEx ensures that a medication mentioned in a sentence like 'patient denies taking aspirin' is correctly attributed to its source but marked as negated rather than active, preventing false-positive reconciliation entries.

The preprocessing step of parsing a free-text clinical note into logical zones such as 'History of Present Illness', 'Discharge Medications', or 'Allergies'. Source attribution at the section level provides critical context—a medication mentioned in 'Past Medical History' carries different clinical weight than one in 'Current Orders'—enabling downstream systems to apply appropriate validation rules.