Industries / Health / Healthcare / Pharmacy
Checking for errors and interactions
Before a medicine leaves the pharmacy, staff and software check that it is the right drug and dose for the right person and that it will not clash with the patient's other medicines.
7
Milestones achieved, from 1990 to 2018.
3
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1990
Screening written into law
The US Omnibus Budget Reconciliation Act of 1990, enacted on 5 November 1990, told each state's Medicaid program to review drug therapy "before each prescription is filled or delivered." The review must screen for duplicated therapy, drugs that conflict with a disease, drug-drug interactions, wrong dose or duration, allergies and misuse. States were first given until 1 January 1993.
Credit
US Congress
2001
Capital letters for look-alikes
Tall man lettering writes the differing part of two similar drug names in capitals, as in vinBLAStine and vinCRIStine. The FDA began its name differentiation project in 2001 and asks makers to adopt the lettering voluntarily when reports show a pair is being confused. A study of 1,676,700 stays at 42 US children's hospitals from 2004 to 2012 found no change in potential look-alike errors after the lettering came into use in 2007.
LessonA cheap fix that looks sensible still has to be measured.
Credit
US Food and Drug Administration; Institute for Safe Medication Practices
2006
Most alerts are overridden
A review of 17 studies found that clinicians override 49% to 96% of the drug safety alerts their ordering software shows them. The authors noted that overriding is often justified, and traced the problem to alerts with low specificity, unclear content and needless interruptions.
LessonThe review put the fault in the software and the organization, not in careless users.
Credit
H. van der Sijs and colleagues, Erasmus University Medical Center, Rotterdam
2006
Scan every dose
An FDA rule of February 2004 required a barcode carrying the National Drug Code, the product's identifying number, on the labels of certain drugs. A 735-bed Boston hospital pharmacy then added barcode scanning to dispensing, and observers checked 115,164 doses before and 253,984 after. Where staff scanned every dose, the dispensing errors that scanning targets fell by 93% to 96%. Where they scanned one dose of a batch, potentially harmful errors rose 2.4-fold. In 2010 the same group reported that scanning at the bedside cut errors in giving medicines from 11.5% to 6.8% of doses.
LessonThe scan protects only the dose that is scanned.
Credit
E. G. Poon and colleagues, Brigham and Women's Hospital
2016
Reporters test 255 pharmacies
Chicago Tribune reporters took prescriptions for five drug pairs with well-established dangerous interactions to 255 pharmacies, most of them in the Chicago area. 52% sold both drugs without mentioning the interaction. Independent pharmacies failed 72% of their tests, CVS 63% and Walgreens 30%. CVS, Walgreens and Wal-Mart then promised changes covering 22,000 stores. The paper called it the largest study of its kind.
LessonA check that exists in the software can still fail at the counter.
Credit
Sam Roe, Karisa King and Ray Long, Chicago Tribune; pharmacy professors Daniel Malone and John Horn chose the drug pairs
2017
A cabinet override kills
An automated dispensing cabinet is a locked, computerized store of medicines on a hospital ward. At Vanderbilt University Medical Center in Nashville, a nurse was to give a patient Versed, a sedative. Prosecutors said she switched the cabinet into override mode, which releases drugs not prescribed to that patient, and took vecuronium, a paralyzing drug. The patient, Charlene Murphey, 75, died. The nurse, RaDonda Vaught, was convicted of criminally negligent homicide in March 2022 and sentenced to three years' probation on 13 May 2022. Other nurses told reporters that overrides are routine in many hospitals.
LessonA bypass that is used every day stops being a safeguard.
Credit
reporting by Brett Kelman, KFF Health News
2018
A defense for honest errors
In the UK a pharmacist who dispensed the wrong medicine by accident could be prosecuted under the Medicines Act 1968 and jailed for up to two years. An Order made on 8 February 2018 created a legal defense for inadvertent errors by registered staff in registered pharmacies. The government estimated that 20,820 dispensing errors were reported in the UK in 2016 and about 499,651 were not, and put a fifth of that gap down to fear of prosecution.
LessonPeople report errors more readily when a report cannot send them to prison.
Credit
UK Department of Health and Social Care
Watch for
AI matches pharmacists on review
RxBench is a set of 2,259 test questions covering 14 common types of prescription error. Its authors ran 18 AI language models on it and report that the leading ones, Gemini 2.5 Pro, Grok 4 and DeepSeek R1, match or beat licensed pharmacists on some tasks.
Achieved whenan outside group reproduces the scores on prescriptions the models have not seen, or a working pharmacy reports fewer missed errors with a model in use.
Claimed by
Yan Yang and colleagues, with authors at Sun Yat-sen Memorial Hospital, November 2025
AI help cuts both ways
In a randomized simulation, 77 pharmacists, residents and students judged medication orders with and without a "verify" or "do not verify" recommendation from OpenAI's GPT-4o. Help raised overall accuracy (adjusted odds ratio 1.58). When the AI was right, users were right 80.2% of the time. When it was wrong, they were right 26.1% of the time, and they accepted 73.9% of its wrong recommendations.
Achieved whena trial on live orders in a working pharmacy shows fewer errors reaching patients, with the harm from wrong AI advice counted.
Claimed by
D. Surdilla and colleagues, UCSF Health, September 2026
Reading controlled-drug prescriptions
A hospital-built system reads images of inpatient prescriptions for narcotic and psychotropic drugs, turns the text into structured data with a locally run language model, and applies a rule base for dose limits, instructions and signatures. On 500 past prescriptions it agreed with pharmacists' manual review 95.6% of the time, with 6 problems missed and 16 false alarms.
Achieved whenthe system is tested on new prescriptions at a hospital that did not build it.
Claimed by
R. Ma and colleagues, hospitals in Qingdao, China, September 2026
Open
01
Sound-alike names, spoken
A public test of speech recognition on spoken prescriptions. Take name pairs from the ISMP List of Confused Drug Names, such as hydralazine and hydroxyzine, write short made-up phone-in orders for invented patients, record them in many voices and accents, and count how often each speech model writes the other drug of the pair.
The briefCredit · unclaimed02
A map of look-alike labels
A ranked list of pairs of different medicines whose package labels look alike, computed from the label pictures in DailyMed, the US National Library of Medicine's label database. An image model turns each label picture into numbers that describe how it looks, and the pairs of different drugs or strengths with the closest numbers go to the top.
The briefCredit · unclaimed
Open 01 · the brief
Sound-alike names, spoken
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test of speech recognition on spoken prescriptions. Take name pairs from the ISMP List of Confused Drug Names, such as hydralazine and hydroxyzine, write short made-up phone-in orders for invented patients, record them in many voices and accents, and count how often each speech model writes the other drug of the pair.
Why it is still open
In August 2024 the Institute for Safe Medication Practices told each hospital to test speech recognition itself, with sound-alike pairs such as clonazePAM and cloBAZam, before using it for medication orders; no shared test exists for them to use. Vendors report their own drug-name figures, such as AssemblyAI's 3.2% rate of missed medical terms on its in-house benchmark, without releasing the test or counting swaps within pairs. Medical speech tests score all clinical terms together. A 2024 study of several speech models on clinical English in 93 African accents found that errors on terms such as drug names ran higher than overall error rates, but it did not isolate confusable pairs. A PubMed search for look-alike, sound-alike names with speech recognition returned nothing.
Built with what exists
ISMP's list, updated through June 2024, and the FDA's tall man lettering list; OpenAI's open Whisper model run locally through whisper.cpp (see the Note "A first stable number for local AI"); a commercial model such as ElevenLabs Scribe v3 (see the Note "Fewer word errors"); text-to-speech voices and volunteer recordings; a scoring script that counts swaps to the paired name separately from other mistakes.
First step
Pick 20 pairs from the list, record each in one sentence with three voices, and run the clips through two models.
Done looks like
At least 100 name pairs, each spoken by at least 10 voices in three conditions (quiet room, phone line, background noise). For each model, a published table of how often it wrote the paired drug, with the audio, reference text and transcripts released.
Open 02 · the brief
A map of look-alike labels
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A ranked list of pairs of different medicines whose package labels look alike, computed from the label pictures in DailyMed, the US National Library of Medicine's label database. An image model turns each label picture into numbers that describe how it looks, and the pairs of different drugs or strengths with the closest numbers go to the top.
Why it is still open
Published work identifies which drug a photo shows; it does not search for confusable pairs. DLI-IT (2020) matched 300 photos of opioid labels to a reference set with up to 88% precision, and a Taiwan hospital trained a model to tell 250 kinds of blister pack apart. Two web searches and a PubMed search found no public ranking of look-alike labels.
Built with what exists
DailyMed's bulk label downloads, whose zip files hold each label's JPEG pictures; an open image model such as DINOv2 or SigLIP; ISMP's confused-names list as a set of known problem pairs; a web page that shows each pair side by side.
First step
Download one monthly DailyMed update (October 2026 held 4,131 labels), pull the label pictures for injectable products, and compute the first similarity scores.
Done looks like
Every pair of labels in one product class, starting with injectable vials, scored. The top 500 pairs published with pictures. Two pharmacists rate a random 100 of them as confusable or not, and the share they both accept is reported.