Can AI Make Medication Safer?
Electronic prescribing with built-in checks has strong Australian evidence behind it: in one study it cut procedural prescribing errors by more than 90%. But every check speaks through an alert, and studies find clinicians override many of them, in some cases nearly all. Machine learning promises fewer, sharper alerts. The evidence for that is promising but early, and it measures alerts, not patient harm.
On this page
Better: Note and confidence
Moderate confidenceFewer prescribing errors, plus new system errorsElectronic prescribing cut potentially serious prescribing errors by 44% on the intervention wards of an Australian study, but errors caused by the system itself made up about a third of the errors that remained.Faster: Note and confidence
Low confidenceChecks are instant; reading them is notWe found no strong evidence on time saved. In our reading, the checks themselves are instant, and the cost is paid in interruptions; in one Australian study fewer than 20% of alerts were read.Cheaper: Note and confidence
Low confidenceThe harm is costly; the savings are modelledA peer-reviewed review estimates medication-related admissions cost Australia at least $1.4 billion a year. The only cost estimate we found for machine-learning alerts was a retrospective US model, not a measured saving.Safer: Note and confidence
Governance-gatedSafer only if alerts stay few and relevantElectronic prescribing cut potentially serious errors in an Australian study, but a 2006 review found drug safety alerts were overridden in 49% to 96% of cases. In our reading, an alert nobody reads protects no one.The problem, in numbers
Medicines cause a large share of the harm that healthcare could prevent.Source 8 The Pharmaceutical Society of Australia estimates that 250,000 Australians are hospitalised each year because of problems with their medicines, and another 400,000 present to emergency departments, with at least half of that harm preventable.Source 1 A peer-reviewed review of Australian studies from 1988 to 2021 put medication-related admissions at a median of about 2.5% of all admissions, about 9% of emergency admissions and about 20.5% of geriatric admissions through emergency. That is at least 250,000 admissions a year (admissions, not people), at an estimated cost of $1.4 billion, a figure its authors call likely conservative.Source 1
The problem is global. The World Health Organization estimates that half of all preventable harm in medical care is medication related, a quarter of it severe or life-threatening, and puts the cost of medication errors at an estimated US$42 billion a year.Source 8 Its Medication Without Harm challenge, launched in 2017, set a goal of halving severe avoidable medication-related harm globally within five years.Source 8 Australia's response was developed by the Australian Commission on Safety and Quality in Health Care.Source 8
What improves, and why
In the Australian studies reviewed here, medication safety software is not machine learning. It is electronic prescribing with rules: a structured order replaces handwriting, and the system checks orders against rules such as drug allergies and drug-drug interactions.Source 3 In our reading, that design removes whole classes of error at once rather than one mistake at a time.
Australian evidence shows how large that effect can be. In a before and after study at two Sydney teaching hospitals, covering 3,291 admissions, total prescribing errors fell by 66.1%, 57.5% and 60.5% on the three wards that adopted electronic prescribing, with no significant change on control wards.Source 2 Procedural errors fell by 90.2% on one hospital's intervention ward and 93.6% at the other hospital.Source 2 Errors rated as potentially serious fell from 0.25 to 0.14 per admission, a 44% reduction. The study counted errors, not harm to patients.Source 2
Machine learning enters at the next step. Rule-based checks can identify only the errors their rules were written for.Source 5 Outlier-detection systems instead flag statistically unusual orders.Source 6 In a retrospective US study of MedAware, a commercial system, 79.7% of a sample of its alerts were clinically valid, and 68.2% would not have been raised by the existing rule-based checks.Source 5
A medication check does not protect a patient. A clinician who reads the alert and changes the order does. Everything between those two points is the safety case.
What the evidence actually says
Electronic prescribing also creates errors of its own. In the Sydney study, errors caused by the system itself, such as choosing the wrong drug from a drop-down list beside the intended one, made up 34.8% of the errors that remained after implementation.Source 2 The authors judged these potentially remediable by redesign and training.Source 2 Clinical error rates, as distinct from procedural ones, did not change significantly overall.Source 2
The bigger problem is the alert. A 2006 review of 17 papers reported that clinicians override drug safety alerts in 49% to 96% of cases, and named the most important reason as alert fatigue: too many alerts that are not serious, not relevant or repeated.Source 4 The same review notes that overriding may often be justified. In our reading, that is the point: when many alerts can safely be ignored, the important one risks being ignored too.
Australian research points the same way. Writing in the Medical Journal of Australia, two researchers cite a ward-round study of Australian doctors in which nearly half of medication orders generated an alert, fewer than 20% of alerts were read and no orders were changed.Source 3 They also cite a review in which drug-condition alerts had positive outcomes in five of six studies, but drug-drug interaction alerts in only two of six.Source 3 They report that all 26 Australian hospitals in one survey had interaction alerts, yet in the four studies that assessed harm, only about 2% of drug-drug interactions were associated with any patient harm, generally of low severity.Source 3
The machine-learning studies aim at this weakness. In a prospective 16-month study on one medical ward of an Israeli tertiary hospital, MedAware, added to the existing alerts, flagged 0.4% of medication orders; 85% of its alerts were clinically valid and 43% led to a change in later orders.Source 6 Sixty percent were raised after the drug was already dispensed, following changes in the patient's status such as vital signs.Source 6 In our reading, checks run only at ordering would miss those. But both studies are small, one retrospective and one on a single ward, and neither measured patient harm.Source 5,Source 6
Where it falls short
Alert fatigue is the failure mode. When checks fire often for things that are not serious or relevant, clinicians can stop heeding them, including the one that matters.Source 4
New errors arrive with the system. About a third of remaining errors in the Sydney study came from the software itself.Source 2
Weak outcome evidence for machine learning. The studies show valid alerts, not fewer injured patients, and the only cost figure we found was a modelled US estimate of more than US$60 of potentially prevented adverse-event costs per alert.Source 5
Where governance decides the outcome
In Australia, clinical decision support software is exempt under the medical device regulations only if it meets all three of the Therapeutic Goods Administration's criteria: it is intended solely to provide or support a recommendation to a health professional, is not intended to directly process or analyse a medical image or signal from another medical device, and is not intended to replace clinical judgement.Source 7 Sponsors must still notify the TGA within 30 days of importation or supply.Source 7 In our reading, that keeps the prescriber accountable, so the hospital must make alerts worth reading. Someone has to own the alert library: which alerts interrupt, which are passive, how often each is overridden and whether that override was right. The MJA authors call for rigorous governance, evaluation and monitoring of alert effectiveness, and note that fewer alerts are likely to work better.Source 3 A machine-learning layer needs the same local validation before it goes live, and the same monitoring after.
How to think about adoption
In our reading, electronic prescribing with core checks is the foundation, and the Australian evidence supports it. Before adding anything, measure what you already have: alert volume per order, override rates by alert type, and how often an override was followed by harm. Switch off or demote alerts that are almost always overridden for good reason, and keep interrupting alerts for the high-risk drugs and conditions where they change decisions. Treat a machine-learning alerting tool as a candidate for that short list, not as an extra layer on top: test it on your own records first, compare its alerts with your existing rules, and judge it by orders changed and harm avoided, not by accuracy alone.
BFCS.ai does not fabricate figures. Rule-based electronic prescribing and machine-learning alerting are kept separate here because they rest on different evidence. The two machine-learning studies cited are small. One is retrospective and was read in abstract only; the other is prospective, on a single ward, and was read in full. For the retrospective study we report only what its abstract states. Neither measured patient harm.
Sources
- 1
Australian burden of medication-related harm. psa.org.au
Lim, R., Kalisch Ellett, L. M., Semple, S. & Roughead, E. E., "The Extent of Medication-Related Hospital Admissions in Australia: A Review from 1988 to 2021," Drug Safety 45(3), 2022, pp. 249-257 (peer-reviewed). pmc.ncbi.nlm.nih.gov
Back to textWhat we took from it
Pharmaceutical Society of Australia, "Medicine Safety: Take Care" (report page): 250,000 Australians are hospitalised each year and another 400,000 present to emergency departments as a result of medicine problems; at least half of this harm could have been prevented.
Drawing on 17 chart-review studies and 8 administrative-data studies: a median of about 2.5% of all admissions, about 9% of emergency admissions and about 20.5% of geriatric admissions via the emergency department are medication related; at least 250,000 hospital admissions a year (admissions, not individual patients, against about 11 million Australian admissions in 2019 to 2020), costing an estimated AUD 1.4 billion (at the average AUD 5,500 per hospital stay), which the authors call likely conservative. The PSA figure counts people hospitalised; the review counts admissions.
- 2
Westbrook, J. I., Reckmann, M., Li, L. et al., "Effects of Two Commercial Electronic Prescribing Systems on Prescribing Error Rates in Hospital In-Patients: A Before and After Study," PLoS Medicine 9(1): journals.plos.org
Back to textWhat we took from it
e1001164, 2012 (peer-reviewed). Two Sydney teaching hospitals, six wards (three intervention, three control, all controls at Hospital A); data collected 2006 to 2010; 3,291 admissions reviewed (1,923 before, 1,368 after). Total errors fell 66.1%, 57.5% and 60.5% on the three intervention wards with no significant change on control wards; procedural errors fell 90.2% (Hospital A intervention ward) and 93.6% (Hospital B intervention wards); serious errors (rated 3 or more on a five-point scale of actual or potential severity) fell from 0.25 to 0.14 per admission (44%) across the intervention wards; system-related errors were 358 of 1,029 post-implementation errors on the intervention wards (34.8%), 99% of them clinical errors; clinical error rates did not change significantly overall. The study counted errors, not measured patient harm. Hospital B had no control ward and wards were not randomised.
- 3
Westbrook, J. I. & Baysari, M. T., "Nudging hospitals towards evidence-based decision support for medication management," Medical Journal of Australia 210(6 Suppl), 2019, pp. S22-S24 (commissioned, externally peer reviewed). mja.com.au
Back to textWhat we took from it
Second-hand figures attributed in the article: a survey of 26 Australian hospitals with electronic medication management found on average five alert types, and all had drug allergy and drug-drug interaction alerts; a review found positive outcomes in five of six studies of drug-condition alerts but two of six studies of drug-drug interaction alerts; in an observational ward-round study of Australian doctors nearly 50% of orders generated an alert, fewer than 20% were read and no orders were changed; about 2% of drug-drug interactions were associated with harm, generally of low severity, in the four studies that assessed it; alert fatigue "is now recognised as a widespread and persistent problem"; the authors call for rigorous governance, evaluation and monitoring of effectiveness and write that fewer alerts are likely to be more effective. Competing interest declared: one author is on the board of the Australian Digital Health Agency.
- 4
van der Sijs, H., Aarts, J., Vulto, A. & Berg, M., "Overriding of drug safety alerts in computerized physician order entry," Journal of the American Medical Informatics Association 13(2), 2006, pp. 138-147 (peer-reviewed literature review). pmc.ncbi.nlm.nih.gov
Back to textWhat we took from it
17 papers included, nine with quantitative override data, split between hospital and outpatient settings and mostly from the US; drug safety alerts are overridden by clinicians in 49% to 96% of cases; alert overriding may often be justified; alert fatigue from a poor signal-to-noise ratio (alerts that are not serious, irrelevant or repeated) is identified as the most important reason for overriding. The review is from 2006 and override rates in current systems may differ.
- 5
Rozenblum, R., Rodriguez-Monguio, R., Volk, L. A. et al., "Using a Machine Learning System to Identify and Prevent Medication Prescribing Errors: A Clinical and Cost Analysis Evaluation," Joint Commission Journal on Quality and Patient Safety 46(1), 2020, pp. 3-10 (peer-reviewed; abstract only read, via PubMed; the publisher page refused automated access). pubmed.ncbi.nlm.nih.gov
Back to textWhat we took from it
MedAware is a commercial system. Alerts were generated retrospectively by a machine learning system (MedAware) on outpatient records from two US academic medical centres, 2009 to 2013: 10,668 alerts; of 300 randomly sampled alerts, 92% were accurate and 79.7% clinically valid; 68.2% would not have been generated by the existing rule-based decision support; estimated cost of potentially prevented adverse events more than USD 60 per alert, about USD 1.3 million when extrapolated to the full patient population; these are modelled estimates, not measured outcomes. PubMed lists no author affiliations and no conflict-of-interest statement, and indexes the work as "Research Support, Non-U.S. Gov't" without naming the funder. We could not read the full text or its funding and disclosure statements, so we cannot confirm whether the vendor was involved.
- 6
Segal, G., Segev, A., Brom, A. et al., "Reducing drug prescription errors and adverse drug events by application of a probabilistic, machine-learning based clinical decision support system in an inpatient setting," Journal of the American Medical Informatics Association 26(12), 2019, pp. 1560-1565 (peer-reviewed; full text read via PubMed Central, the publisher page being subscriber-only). pmc.ncbi.nlm.nih.gov
Back to textWhat we took from it
Prospective study of MedAware, a commercial system, added to an existing alerting system on one medical ward of Chaim Sheba Medical Center, Israel, monitoring all prescriptions over 16 months: 315 alerts on 282 of 78,017 medication orders (0.4%); 85% clinically valid and 80% clinically useful; 43% caused changes in subsequent orders; 60% were flagged after the medication was already dispensed, following changes in the patient's status such as vital signs. All six authors list Sheba Medical Center and Tel Aviv University affiliations; the paper declares no specific funding and no competing interests. Patient outcomes were not measured, despite the title's reference to adverse drug events.
- 7
Therapeutic Goods Administration, "Notification form: Clinical decision support software exemption" (opened 4 January 2021), under the Therapeutic Goods (Medical Devices) Regulations 2002. consultations.tga.gov.au
Back to textWhat we took from it
Software is exempt only if all three criteria are met: it is intended solely to provide or support a recommendation to a health professional; it is not intended to directly process or analyse a medical image or signal from another medical device; and it is not intended to replace the clinical judgement of a health professional. Sponsors must notify the TGA within 30 days of importation or supply.
- 8
World Health Organization. who.int
WHO news release, "WHO calls for urgent action by countries for achieving Medication Without Harm," 16 September 2022: who.int
Australian Commission on Safety and Quality in Health Care, "WHO Global Patient Safety Challenge" (page last updated 30 April 2026): safetyandquality.gov.au
Back to textWhat we took from it
"Medication Without Harm" initiative page: launched at the Second Global Ministerial Patient Safety Summit, Bonn, 29 March 2017; aims to reduce severe avoidable medication-related harm by 50% globally within five years; the cost of medication errors has been estimated at USD 42 billion a year.
half of all preventable harm in medical care is medication related, and a quarter of that is severe or life-threatening.
the Commission was engaged to develop Australia's response to the challenge, which includes recommendations and priority action areas to reduce medication errors, adverse drug events and medication-related harm; a status report uses data to June 2023.