Medication Safety & Prescribing

AI applied to the medication itself: catching a prescribing error before it reaches the patient, computing a dose from that patient's own pharmacokinetics, screening a regimen for interactions and cumulative risk, scoring a patient for controlled substance risk, and surveilling an inpatient population for adverse drug events and antimicrobial misuse. The category separates from clinical decision support on what the model decides. Here the decision is about a drug, which drug, what dose, and whether this patient should receive it, and the output is either administered or used to withhold something. That distinction has a regulatory expression as well as an editorial one. Several products in this category are cleared by the FDA as Class II software precisely because they compute a dose, while others assert the Cures Act exclusion for non device clinical decision support on the ground that a clinician can independently review the basis of their output, a claim that is currently the subject of a citizen petition asking the agency to decide the question. Establishing which side of that line a product sits on, and on whose reading, is the first thing to ask a vendor here and it is rarely on the datasheet. Three failure modes define the category and they are not the same failure. The first is the wrong quantity. A dose produced by a pharmacokinetic model fitted to two blood levels cannot be reproduced in a clinician's head, so the ordinary safeguard of independent review is unavailable exactly where the consequence is kidney injury or hypoglycaemia. The second is the ignored alert. This is the segment that produced the override problem in the first place, and legacy interaction checking is routinely overridden at rates that would be treated as a defect anywhere else, so the useful metric is not sensitivity but alert burden per order and the proportion of alerts that change a prescription. The third is the withheld medication, and it exists nowhere else in this index. A risk score attached to a patient's controlled substance history can result in that patient being refused treatment, so a false positive here is not wasted time but denied care, and the published false positive rate for the most widely deployed such score is high enough that roughly a third of scored patients may be misclassified. Buyers should ask four things. What is the alert burden per order and what proportion of alerts result in a changed prescription, since a vendor who cannot answer is selling into the problem it claims to solve. Whether the model is the vendor's own or licensed, because several incumbents in this category embed another vendor's dosing engine and sell it as a feature. Whether the algorithm and its validation are published in a form an outside party can check, since the most consequential products here are also the most opaque. And whether performance has been reported by subgroup, which matters more than usual in a category where the scored behaviors, pharmacy counts, prescriber counts and distance traveled, track access to care at least as well as they track risk.