Medication Reconciliation (MedRec)

A rigorous MedRec process categorizes every variance between medication lists using standardized taxonomies. This moves beyond simple mismatches to clinically meaningful classification.

• Unintentional Discrepancy: An unjustified difference without clinical rationale, representing a medication error.
• Omission Error: A clinically indicated drug unintentionally not prescribed on new orders.
• Commission Error: A drug added without a valid indication.
• Duplication Error: Co-prescribing therapeutically equivalent agents.
• Dosage/Strength Error: A mismatch in dose, frequency, or route. This structured approach enables targeted quality improvement and system-level error proofing.

The foundation of MedRec is the Best Possible Medication History (BPMH) , which is not a simple copy of the last electronic health record list. It requires a systematic interview and verification against at least two independent information sources.

• Primary Sources: Patient/caregiver interview, physical medication bottles.
• Secondary Sources: Community pharmacy dispense records, insurance claims data, previous discharge summaries.
• Tertiary Sources: Pill bottle inspection, blister pack review. Discrepancies between sources must be explicitly resolved and documented to create a single, trusted pre-admission list.

MedRec is not a static snapshot but a chronological reconstruction of a patient's medication journey. Temporal reasoning is the capability to sequence clinical events to validate logical consistency.

• Validates that a medication's start date precedes its discontinuation date.
• Identifies therapeutic overlaps or gaps in therapy during transitions.
• Reconciles 'as needed' (PRN) medications with administration records to determine actual active status.
• Detects unintentional duplicate therapy by analyzing concurrent active order windows. This time-aware analysis is critical for distinguishing intentional, clinically appropriate changes from accidental errors.

To compare medication lists from disparate sources, a MedRec system must computationally normalize heterogeneous representations into a common, comparable format.

• Active Ingredient Matching: Links brand names (e.g., Tylenol) and generics (e.g., acetaminophen) to a base RxNorm concept identifier to prevent duplicate therapy errors.
• Dose Normalization: Converts varied expressions of strength (e.g., '500 mg', '1 tablet', 'two 250 mg capsules') into a standardized cumulative daily dose for accurate comparison.
• Frequency Standardization: Maps vernacular instructions like 'take every morning' to coded, computable frequencies (e.g., 'QAM', 'Once Daily'). Without this semantic normalization, automated comparison is impossible.

Identification of a discrepancy is only the first step. A complete MedRec process requires a closed-loop resolution workflow that drives action and documents clinical intent.

• Intentional Discrepancy: The clinician documents the clinical rationale for a deliberate change (e.g., 'held due to acute kidney injury').
• Unintentional Discrepancy: The error is corrected, and the order is updated in the active medication list.
• Documentation: The final reconciled list is clearly documented as the single source of truth for the next care provider. This loop ensures that the reconciliation is not just an audit but an intervention that corrects the patient's active orders.

In high-stakes clinical environments, fully autonomous reconciliation is unsafe. A Human-in-the-Loop (HITL) design paradigm is essential, where AI outputs are treated as decision support for a qualified clinician.

• Confidence Thresholding: AI-extracted data with a low model prediction score is routed for mandatory pharmacist review.
• Source Attribution: Every extracted medication entry is explicitly linked back to its source sentence or database field, enabling rapid human verification.
• Audit Interface: A specialized UI allows a clinical pharmacist to efficiently accept, reject, or modify AI-proposed discrepancies. This architecture optimizes the balance between automation efficiency and patient safety, ensuring a licensed professional always validates the final record.

The core logic of MedRec is classifying differences between a patient's pre-admission medication list and new orders. An unintentional discrepancy is an unjustified difference without clinical rationale—a medication error requiring immediate resolution. In contrast, an intentional change is a deliberate, clinically justified action (e.g., holding an anticoagulant before surgery). AI systems must distinguish between these two categories by analyzing structured order reasons and free-text clinical notes to avoid flagging appropriate therapeutic decisions as errors.

The most common and insidious type of medication discrepancy. An omission error occurs when a clinically indicated drug that a patient was taking prior to a care transition is unintentionally not prescribed on the new medication orders. This is particularly dangerous for time-critical medications such as antiplatelets, anticonvulsants, or Parkinson's disease therapies, where even a single missed dose can precipitate a significant adverse event. Automated reconciliation engines prioritize flagging these silent gaps.

A clinical decision support notification triggered when a new medication order is placed for a drug that is therapeutically equivalent or identical to an existing active order. This poses a direct risk of overdose. Advanced systems use active ingredient matching to resolve brand-name and generic products to a common base compound, preventing duplicate therapy errors that arise from proprietary naming conventions. The challenge lies in tuning alert sensitivity to avoid contributing to alert fatigue.

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 human-in-the-loop (HITL) paradigm optimizes the balance between straight-through processing and clinical safety. High-confidence extractions are auto-committed, while ambiguous cases—such as illegible handwriting or rare drug names—are surfaced to a clinical pharmacist for verification, ensuring that automation never outpaces accountability.

Deterministic constraints and post-processing rules applied to large language model outputs to prevent the generation of plausible-sounding but factually non-existent medication names or dosages. These guardrails include validation against authoritative drug databases like RxNorm, regex-based dose format enforcement, and rejection of any generated entity not grounded in the source text. In the high-stakes domain of medication reconciliation, a hallucinated 'metformin 1000mg' that should be 'metoprolol 100mg' is a catastrophic failure mode that must be engineered out of existence.