Prescription Entered Into the Dispensing System
Every dose, unit, and frequency starts as raw entered text.
- 420: Median entries per pharmacy/day (prescriptions processed)
- 1 in 200: Manual entry error rate (before AI review)
- ~15%: Decimal-point error share (of dosing errors)
- 6: Fields captured per order (dose, unit, freq, route, weight, form)
Structured capture of the order
Dose, unit, frequency, and route are parsed into fields.
Why entry is the highest-risk step
Most dosing errors originate at transcription, not calculation.
Comparing Against Known Dosing-Error Patterns
The AI checks the order against a curated library of past errors.
- 1,200+: Error patterns in library (curated signatures)
- 3: Pattern categories (decimal, unit, weight-based)
- Weekly: Library update cadence (new cases added)
- <80ms: Lookup latency (per order)
Signature matching approach
Each order is embedded and compared to known error vectors.
False positive control
Only high-similarity matches surface for review.
Scanning for 10× Decimal Shifts and Unit Confusion
The model tests whether the dose looks like a shifted decimal or swapped unit.
- 10×: Common decimal shift (most frequent multiple)
- mg ↔ mcg: Unit-pair confusions (most frequent swap)
- 94%: Detection sensitivity (on labeled test set)
- 0.3s: Median scan time (per prescription)
Decimal-shift detection
Dose is checked against 0.1× and 10× neighbors of the norm.
Unit confusion detection
mg, mcg, and mL are cross-checked against drug defaults.
Verifying Pediatric Dose Against Patient Weight
Weight-based doses are recalculated and compared to the order.
- ~80%: Pediatric orders weight-based (of dosing)
- ±10%: Typical mg/kg tolerance (accepted variance)
- 1 in 500: Weight-mismatch error rate (orders flagged)
- <1s: Recalculation time (per order)
Recomputing the mg/kg dose
Entered dose divided by recorded weight, compared to norms.
Stale-weight risk
Outdated weight entries silently distort safe dosing.
Matching Error Pattern Identified and Flagged
A confident match routes the order to a pharmacist before dispensing.
- 85%: Auto-flag confidence threshold (triggers hard stop)
- ~40%: Pharmacist review time saved (vs manual triage)
- 1 in 3,000: Reported error interceptions (orders processed)
- <1s: Time to flag (end-to-end)
Routing the flagged order
High-confidence matches are held from dispensing automatically.
Human-in-the-loop confirmation
A pharmacist confirms or overrides before release.