Medication Extraction

Identifies the active ingredient or brand name along with its numeric strength and unit of measure. This foundational attribute distinguishes between formulations like 'Lisinopril 10 mg' and 'Lisinopril 20 mg', which have different therapeutic implications.

• Brand vs. Generic: Maps trade names like 'Zestril' to their generic equivalent 'Lisinopril'
• Compound Medications: Detects multi-ingredient drugs such as 'Lisinopril/Hydrochlorothiazide 20/12.5 mg'
• Normalization: Links extracted strings to standardized RxNorm concept unique identifiers for interoperability

Extracts the quantitative amount to be taken and the temporal schedule governing administration. This pair is critical for calculating daily intake and identifying potential overdose scenarios.

• Dosage: Captures structured values like '2 tablets' or '10 mL'
• Frequency: Resolves abbreviations such as 'BID' (twice daily), 'TID' (three times daily), and 'QHS' (at bedtime)
• Complex Schedules: Handles tapering instructions like 'Take 2 tablets daily for 7 days, then 1 tablet daily'

Classifies the anatomical path by which the medication enters the body. The route directly impacts bioavailability and is essential for distinguishing safe from erroneous orders.

• Common Routes: Oral (PO), Intravenous (IV), Intramuscular (IM), Subcutaneous (SC), Topical
• Ambiguity Resolution: Differentiates 'PO' (by mouth) from 'PO' (post-operative) using contextual embeddings
• Safety Implications: Flags mismatches like an IV-only formulation ordered with an oral route

Captures the total length of therapy and the dispensed amount, enabling calculation of days' supply and identification of truncated or extended regimens.

• Duration: Extracts explicit statements like 'for 10 days' or 'until follow-up'
• Quantity: Parses dispense amounts such as '#30' or 'Qty: 60 mL'
• PRN Designation: Identifies 'as needed' medications where duration is indefinite and tied to symptom recurrence

Reconstructs the full signatura—the patient-facing administration instruction—as a unified, structured object. The sig synthesizes all extracted attributes into a single, actionable directive.

• Free-Text Parsing: Converts 'Take 1 tab by mouth twice daily for 14 days' into discrete structured fields
• Latin Abbreviation Expansion: Translates archaic shorthand like '1 tab PO BID x 14d' into plain English
• Standardization: Outputs a canonical representation suitable for pharmacy system ingestion and patient education materials

Determines whether the medication mention refers to a current active order, a historical prescription, or a discontinued therapy. This temporal grounding prevents double-counting and ensures accurate medication reconciliation.

• Status Classification: Labels mentions as Active, Discontinued, Historical, or Planned
• Date Association: Links the medication to specific start and stop dates when explicitly stated
• Section Awareness: Leverages document zones like 'Discharge Medications' vs. 'Home Medications' to infer status when dates are absent

The foundational annotation scheme for training medication extraction models. Tokens are labeled as Beginning, Inside, or Outside of an entity span.

• B-DRUG: Marks the first token of a medication name
• I-DOSAGE: Continues a dosage entity like '500 mg'
• O: Tokens outside any entity

This scheme enables sequence labeling models to learn precise entity boundaries, distinguishing 'Aspirin 81 mg daily' as three distinct but related entities.

The critical step of mapping extracted medication strings to standardized RxNorm Concept Unique Identifiers (RxCUIs) . This resolves lexical variability into a single semantic identity.

• 'Tylenol 500 mg tablet' maps to RxCUI: 209459
• 'Acetaminophen 500 mg oral tablet' maps to the same RxCUI
• Enables drug-drug interaction checking and allergy cross-referencing

Without normalization, extracted mentions remain ambiguous strings rather than actionable, computable data points.

Determining whether an extracted medication is actually relevant to the patient's active regimen. Contextual algorithms like NegEx and its successors identify:

• Negated: 'Patient denies taking aspirin'
• Historical: 'Previously on warfarin'
• Hypothetical: 'Consider starting metformin'
• Family history: 'Mother takes lisinopril'

This prevents false positives in medication reconciliation, ensuring only current, affirmed medications populate the active list.

Medication mentions exhibit hierarchical structure where attributes are contained within larger entities. Nested NER handles this complexity directly.

• Outer entity: 'Ibuprofen 600 mg PO q6h PRN pain'
• Inner entities: Dosage (600 mg), Route (PO), Frequency (q6h), Condition (PRN pain)

Standard flat NER models struggle with overlapping spans. Span categorization and parsing-based approaches explicitly model these containment relationships for complete structured extraction.

A high-precision method using curated gazetteers of medication names and attributes. Efficient data structures like prefix trees (tries) enable rapid matching against millions of terms.

• Matches exact strings: 'atorvastatin', 'Lipitor'
• Handles lexical variants via fuzzy matching
• Complements statistical models in hybrid NER systems

Dictionary methods excel at catching rare drugs and branded formulations that may be underrepresented in training data, but require continuous curation as formularies evolve.

Medication extraction is rarely a standalone task. It operates within a modular pipeline architecture that orchestrates sequential processing stages:

• Section splitting: Isolates medication list sections from narrative
• Sentence detection: Handles clinical shorthand and abbreviations
• Tokenization: Splits '500mg' into meaningful subword units
• NER: Identifies drug, dosage, route, frequency
• Normalization: Maps to RxNorm
• Context validation: Filters negated or historical mentions

Tools like cTAKES and MedSpaCy provide pre-built pipelines that integrate these components for production deployment.