Evidence
Not individual ones. The safeguards that should catch an unsafe dose go quiet at the one moment that decides whether it reaches the patient — when a nurse takes the medication from the cabinet. Safer-Med adds a systematic, patient-specific check at exactly that point, where none exists today.
The literature below establishes that these risks are real and reach patients.
Errors around ADC overrides, and orders that turn unsafe as a patient’s physiology changes, reach patients — and existing safeguards do not eliminate them. These are documented failure modes, shown below with their settings and limits.
It does not show that Safer-Med prevents these events. Safer-Med’s incremental detection over existing safeguards, its alert burden, and its clinical benefit are still to be validated with development partners — the purpose of the retrospective and shadow phases.
01 · The burden
In a 2023 New England Journal of Medicine study of 2,809 admissions across 11 Massachusetts hospitals, 23.6% had at least one adverse event. Adverse drug events were the single largest category — 39.0% of all events — and 26.8% of them were judged preventable. Read as: ADEs were 39% of adverse events, not of admissions. A contemporary burden anchor, not a dispense-specific figure.
02 · ADCs and overrides
Safer-Med is not an argument that dispensing cabinets are unsafe. ADCs are a safety technology; the evidence below is about the measurable failure modes that persist around them — the residual gap Safer-Med is built to close.
A 2026 systematic review of 16 hospital studies (2008–2025) found ADC implementation was generally associated with fewer reported medication errors — especially omissions and wrong-drug or wrong-dose events — though error definitions varied and the evidence is mostly observational. Read as: Safer-Med addresses residual risk in a technology that already helps, not a broken one.
At one tertiary hospital, profiled-ADC overrides were held to a 1.9% rate — yet among 664 analyzed overrides, investigators found unrecognized override errors in 5.3%, alongside missing orders and undocumented verbal orders; only 3 (0.5%) met every expected safety element. Read as: a single-site analysis, but it shows override rate alone is a weak proxy for appropriateness.
In a 343-bed pediatric emergency department, among the override pulls evaluated, pharmacist verification occurred a median of four minutes before administration — so many overrides were given at or after the order was already verified, not to save time. Read as: single-site retrospective; bypassing pharmacist review is frequently not the faster path.
Pennsylvania’s reporting system logged 583 technology-override medication-error reports (2013–14); 26.4% involved at least one high-alert medication, and automated dispensing cabinets were the most commonly reported technology. Read as: voluntary reports describe real failure types — not an incidence rate per transaction.
03 · The last mile
Each of these is a systemic failure — the check that should have caught the wrong medication was not there when it left the cabinet.
In the widely reported 2017 case at Vanderbilt University Medical Center, a nurse used an automated-dispensing-cabinet override, searched “VE,” and withdrew the paralytic vecuronium instead of the ordered sedative Versed (midazolam). The patient, Charlene Murphey (75), died. A later peer-reviewed analysis emphasized the system failures — a default that displayed only generic names, an override that surfaced any drug, and no independent check at removal. Read as: a single, widely documented case illustrating the risk of bypassing pharmacy review — a systemic failure, not proof Safer-Med would have prevented it.
An AHRQ Patient Safety Network case (2024): after a verbal order for haloperidol, an override released midazolam instead. The patient was deeply sedated for about two hours and required flumazenil to reverse the benzodiazepine. Read as: a case with several contributing factors, including a verbal order — illustrative of the override risk, not attributable to the cabinet alone.
An ICU nurse reaching for naloxone (Narcan) was thrown by the look-alike name Norcuron and pulled the paralytic vecuronium from the cabinet instead. The patient went into cardiac arrest, was resuscitated and ventilated, and later recovered fully. Read as: an ISMP-reported case, not an incidence study — the same wrong-drug pathway, a different outcome.
At Methodist Hospital in Indianapolis (2006), heparin 10,000 units/mL was loaded where 10 units/mL flush vials were expected; three premature infants died and three more were injured. The identical mix-up at Cedars-Sinai (2007) overdosed actor Dennis Quaid’s newborn twins, who survived after reversal — a case that drew national attention. Read as: a stocking and labeling failure, not an override or a lab-driven event; it shows how last-mile system errors slip past expectation-based checks.
04 · The moving target
Among 396 hospitalized patients whose creatinine shifted while on nephrotoxic or renally cleared drugs (CJASN, 2013), 43% experienced an adverse or potential drug event or therapeutic failure, two-thirds of them preventable. There were 73 actual ADEs — most serious or life-threatening, one fatal, three requiring acute dialysis. Read as: the 43% is a composite that includes potential events; the closest direct support for the “safe earlier, unsafe later” problem.
Across 457 inpatient insulin-associated hypoglycemic events, an indication to reduce the insulin dose was present in a third to nearly half of patients in the day or two beforehand — yet in many, the dose was not substantially reduced. Read as: antecedent risk signals and therapeutic inertia, not proof that a dispense-time alert would have prevented each event.
05 · The noise problem
After a commercial EHR go-live, renal dose-change alerts were judged appropriate only 12.5% of the time, and every alert was overridden across the study year (JAMIA, 2021). Investigators identified five ADEs — four of them followed inappropriately overridden alerts. Read as: poor targeting conditions a workflow in which even the alerts that matter are dismissed.
In a prospective ICU study of 2,448 overridden alerts (BMJ Quality & Safety, 2018), most overrides — 81.6% — were appropriate, but ADEs followed the inappropriate ones far more often: 16.5 versus 2.74 per 100 (adjusted odds ratio 6.14). Read as: clinical-decision-support overrides at one site, not cabinet overrides. The danger is the small wrong subset, not overrides in general.
Clinicians overrode 95.7% of high-priority drug-interaction alerts in a commercial system (JAMIA, 2020) — and among the medication pairs actually given, ADEs were more common after inappropriate than appropriate overrides (9.4% versus 4.3%). Read as: order-entry alerts, not dispense-time. A 95.7% override rate is not 95.7% unsafe behavior — it is what alert noise produces.
Across 112 clinicians (BMC Medical Informatics, 2017), the likelihood of accepting a reminder dropped about 30% for each additional one in the same visit. Read as: outpatient behavioral evidence with no harm endpoint — it explains the mechanism of alert fatigue, but doesn’t quantify inpatient risk.
06 · Signal, not volume
In a randomized trial of 5,060 adults with acute kidney injury (Nature Communications, 2023), medication-targeted alerts increased discontinuation of the risky drug (61.1% versus 55.9%) — but the patient outcome, a composite of AKI progression, dialysis, or death, was not significantly improved. Read as: provider-facing AKI alerts, not dispense-time nurse alerts; a null clinical result despite a real behavior change, not proof targeted signals can’t work.
A six-hospital randomized trial of a generic AKI alert (BMJ, 2021) found no overall outcome benefit, with worse results at two non-teaching sites. Read as: evidence of heterogeneity and a caution against generic alerting — the design lesson is specificity, not volume.
The lesson across these trials is the one the alert-fatigue studies teach: volume, and even clinician response, are not the goal. A precise, patient-specific signal that isn’t lost in noise is — and that is the bar Safer-Med is built to meet.
The approach
Every step above points to the same gap — at the one point in the chain, the cabinet at the moment of dispense, no patient-specific check exists today. Safer-Med adds that check: it re-evaluates the medication against the patient’s current clinical picture, keeps the signal specific so the alert that matters reaches the nurse rather than drowning in noise, and is designed to be measured against adjudicated preventable harm — not alert counts.
References
Statistics are drawn from the cited abstracts, full text, or authoritative source pages. Where evidence is observational, a case report, or a subgroup result, that limit is stated. The literature establishes that these risk pathways are real; it does not establish that Safer-Med prevents these events.