Where does patient context go in a dose check?

The patient context is already there

The detail that gives me most pause in Doximity’s medication-dosing evaluation is the loss of patient context. The company reports that the most common problem was failure to adjust for characteristics such as age, kidney function or drug interactions. Those details were already supplied in the question. Addressing this kind of error therefore involves more than collecting additional information: the test also has to examine how an answer uses information already available to it.^[1]

The company describes testing 9 models released in 2026 with more than 8,000 questions covering 2,204 drugs. At least 2.3 percent of 261,508 evaluated dosages were wrong, a figure Doximity calls a lower bound. The denominator here is dosages; the separate finding of nearly 20 serious errors per 1,000 answers uses answers. Keeping those units visible helps prevent a further mistake: treating either figure as a measured probability of harm to a patient.^[1]

What enters the verification step

This failure mode makes DoseCheck’s design interesting. Doximity says the tool extracts the drug, dose, route, frequency and relevant indication from an answer, then compares them with structured drug information. The same dosage claim returns the same verification result each time. A badge beside a matching dose links to the underlying information. That is a concrete way to make the basis of the check visible: the user can open the source and inspect the details.^[1]

The evaluation chain I want to examine starts before that comparison. If an important patient condition is lost in the extracted information, a consistent comparison leaves that omission unresolved. This is a proposed test, rather than a finding that DoseCheck makes that mistake. The context errors reported by the company explain why the test matters. There is also a plausible alternative: the limiting factor could be insufficiently detailed database coverage for a patient group, rather than the extraction step.^[1]

The clinician beside the badge

Healthcare Dive also reports Doximity’s claim of 99.9 percent accuracy for its own tool, noting that the research accompanied the product launch. An independent publisher reporting the comparison supplies no independent replication of the tool. The developer’s findings still have value as an early warning: a fluent answer containing a dose unsuited to the patient gives verification a concrete target. Which errors the reported accuracy covers is among my first questions when assessing the tool.^[2]

Doximity explicitly says DoseCheck preserves the need for clinical judgment. For a next evaluation, I would want another team to use questions containing patient conditions and measure both missed errors and unnecessary alerts. The comparison should include the time clinicians spend consulting the linked information and whether they correct an important error. That connects the badge’s value to the work facing the clinician, alongside its matching accuracy. A claim of patient benefit requires a further evaluation that follows patient outcomes.^[1]