Multiple Disparate Medication Record Sources
Pharmacy fills, hospital EHR notes, and patient-reported lists rarely agree.
- 3: Source types ingested (pharmacy, EHR, patient)
- 8–15: Avg. records per encounter (across all sources)
- High: Format variety (structured + free text)
- ~35 min: Manual reconciliation time (per patient, pre-automation)
Why sources disagree
Each system captures medications in its own format.
Pharmacy claims lag prescribing by days or weeks.
Structured vs unstructured input
EHR fields are structured; patient recall is free text.
The automation opportunity
An AI engine can normalize all three sources at once.
NLP Extraction of Medication Mentions
A named-entity model pulls drug, dose, and frequency from text.
- 4: Entity types extracted (drug, dose, route, frequency)
- Clinical NER: Extraction model (transformer-based)
- ~97%: Accuracy on clean text (benchmark performance)
- ~78%: Accuracy on messy text (handwritten or OCR input)
Named entity recognition
The model tags spans of text as drug or dose.
Abbreviations like "qd" and "BID" are normalized automatically.
Handling noisy input
OCR errors and shorthand lower extraction confidence.
Confidence scoring
Every extracted entity carries a machine confidence score.
Entity Matching Across Record Sources
The same drug is matched even under different names.
- 4: Matching signals used (name, dose, class, timing)
- 1,200+: Brand/generic pairs known (reference database)
- >95%: Match precision target (to avoid false merges)
- 80–95%: Typical match rate (depends on data quality)
Fuzzy name matching
Brand and generic names are mapped to one concept.
Zestril and Lisinopril resolve to the same entity.
Similarity scoring
Dose and frequency proximity boosts match confidence.
Unmatched entities
Single-source medications stay flagged as unconfirmed.
Dose and Frequency Discrepancy Flagging
Matched entities are compared for conflicting dose or frequency.
- 3: Discrepancy types tracked (dose, frequency, route)
- ~15–25%: Flagged for review (of matched pairs)
- Anticoag, insulin: High-risk drug classes (prioritized review)
- <5%: False-flag rate (after confidence tuning)
Rule-based comparison
Numeric dose values are compared after unit normalization.
A 20mg vs 40mg mismatch triggers an immediate flag.
Risk-weighted prioritization
High-risk drug classes get flagged first for review.
Human-in-the-loop review
Every flag routes to a pharmacist for confirmation.
Reconciled Medication List Generation
One list is assembled from matched and flagged entities.
- 3: Sources merged (into one record)
- ~25 min: Time saved per patient (vs manual review)
- ~70%: Discrepancies resolved pre-visit (auto-triaged)
- <5 min: Clinician review time (per flagged list)
Single source of truth
The reconciled list becomes the record of record.
Flags stay attached until a clinician resolves them.
Audit trail
Every merge and flag keeps its source provenance.
Continuous learning
Confirmed and rejected flags retrain the matching model.