Data Provenance

The mechanism of explicitly linking each extracted medication entry back to the specific sentence, document, or database field from which it was derived. This granular traceability enables rapid human verification by allowing a clinical pharmacist to click on a discrepancy and immediately view the original source text. Without source attribution, AI-extracted medication lists become unverifiable black boxes that undermine clinical trust.

• Links extracted RxNorm codes to original EHR text spans
• Enables single-click audit from output to source
• Critical for HIPAA compliance and medicolegal defensibility

A complete, timestamped record of every computational operation applied to a medication data point as it moves through the reconciliation pipeline. This includes dose normalization, active ingredient matching, and RxNorm mapping steps. Transformation lineage answers the question: 'How did the system arrive at this structured output from the raw input?'

• Captures each mapping from proprietary drug names to RxNorm CUIs
• Records dose unit conversions (e.g., mg to mcg)
• Documents the specific model version and prompt used for extraction

The chronological metadata that timestamps when a medication data point was created, modified, or deprecated across all source systems. Temporal provenance is essential for temporal reasoning—validating that a medication start date logically follows a discontinuation date. It prevents the dangerous scenario where an outdated medication list is used for admission orders.

• Records original observation time vs. system entry time
• Tracks data staleness for real-time reconciliation
• Supports chronological sequencing of clinical events

Explicit identification of the source system that generated each medication record—whether an EHR, pharmacy dispensing system, patient portal, or external health information exchange. This metadata is critical for adjudicating conflicts when two sources disagree about a patient's medication regimen. The provenance record preserves the system identifier, its trust level, and the original data format.

• Distinguishes EHR orders from pharmacy fills from patient-reported data
• Assigns confidence weights based on source reliability
• Preserves original HL7 v2, FHIR, or CDA format reference

The complete log of every human-in-the-loop interaction with an AI-extracted medication record, including who reviewed it, what changes were made, and when the review occurred. This component of provenance ensures that manual overrides and clinical judgments are preserved as part of the permanent record, maintaining accountability across the reconciliation workflow.

• Captures reviewer identity, timestamp, and action taken
• Records the confidence threshold that triggered the review
• Preserves both the original AI output and the human-corrected version

The architectural pattern of storing provenance records in append-only, tamper-evident data structures that prevent retroactive alteration. This ensures that the complete history of a medication data point—from extraction through validation—remains forensically intact for regulatory audits, legal discovery, and patient safety investigations.

• Uses cryptographic hashing for tamper detection
• Supports W3C PROV standard data model for interoperability
• Enables replay of the exact data state at any point in time

The mechanism of explicitly linking each extracted medication entry back to the specific sentence, document, or database field from which it was derived.

• Enables rapid human verification by showing exactly where a data point originated
• Critical for resolving discrepancies during pharmacist review
• Often implemented as span-level annotations in the original clinical text
• Example: Highlighting the exact phrase 'Lisinopril 10mg daily' in a discharge summary that generated a structured medication entry

A consolidated, cross-encounter view of a patient's prescribed, dispensed, and administered medications over time, serving as the single source of truth for the reconciliation engine.

• Aggregates data from EHRs, pharmacy claims, and patient-reported histories
• Each entry carries immutable provenance metadata including source system and timestamp
• Enables temporal reasoning across care transitions
• Foundation for detecting unintentional discrepancies between admission and discharge orders

A Fast Healthcare Interoperability Resources profile that records a patient's reported or derived statement of medication intake, capturing the 'what was taken' aspect of the medication history.

• Includes a derivedFrom element that explicitly references the source resource
• Supports provenance tracking through FHIR's Provenance resource
• Distinguishes between patient-reported, practitioner-reported, and system-derived statements
• Essential for maintaining data lineage when exchanging medication information across systems

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.

• Directly tied to provenance: low-confidence extractions require source verification
• Optimizes the balance between automation and safety in reconciliation workflows
• Typical thresholds set at 0.85–0.95 for medication name extraction
• High-confidence entries may bypass manual review but retain full provenance audit trail

The preprocessing step of parsing a free-text clinical note into logical zones—such as 'History of Present Illness' or 'Discharge Medications'—to increase extraction accuracy.

• Provides structural provenance by tagging which section a medication was extracted from
• Critical context: a medication in 'Allergies' vs. 'Active Orders' has vastly different clinical meaning
• Enables section-specific confidence weighting
• Uses header detection and NLP-based boundary classification

A regular expression-based NLP algorithm designed to identify whether a clinical finding mentioned in narrative text has been negated by the clinician.

• Essential for provenance accuracy: 'patient denies taking aspirin' must not generate an active medication entry
• Detects negation triggers like 'no,' 'denies,' 'without evidence of'
• Operates within a configurable window of tokens before and after the target term
• Foundation for distinguishing affirmed vs. negated medication mentions in source documents