Real-time drug interaction monitoring

The check is done when the alert names both drug orders and cites the interaction monograph those orders map to. If you cannot point at two orders and a monograph row, you do not have an interaction finding. A one-sided flag is a quality miss, not a softer warning.

Hospital pharmacists already run this comparison by hand: a new prescription against what is actually scheduled, then a look-up in the knowledge source, then a conversation with the prescriber. Continuous ML can repeat that comparison as orders change. It does not get to skip the cite, invent a severity the monograph does not state, or turn a flag into a hold.

Patient-specific risk factors belong in the same pass only as facts already on the chart: renal function, age, documented interacting disease, or a genotype another workflow already resolved. Those facts sit next to the cite. They do not replace naming both drugs.

One illustrative pass, not a measured case. A patient has an active MAR row for simvastatin. A new order for clarithromycin posts. The usable alert names the simvastatin order, names the clarithromycin order, and cites the monograph entry for that pair. If the knowledge source returns no matching pair, the alert field stays empty. The prescriber still decides whether to change therapy, adjust the dose, add monitoring, or proceed.

Load the medication administration record as the source of active therapy. Do not substitute a free-text home list, a problem-list mention, or a discharge summary from a prior stay. The MAR is what nursing will administer against. Load the new prescription as the candidate with the same fields. Only then request the monograph for that pair.

medication reconciliation agent answers a different question: is the list complete. This check answers whether two named products interact. Interaction logic on an unreconciled list will miss pairs that never reached the MAR and will flag pairs that were already stopped.

discharge medication reconciliation is adjacent, not a substitute. Discharge work changes the home regimen. Inpatient real-time monitoring watches orders that are about to be given. Do not treat a discharge rec output as the live MAR.

Monograph text typically comes from a drug knowledge base such as FDB. Order context typically lives in the EHR, including systems such as Epic or Oracle Health. Specialty genomic or oncology context may already sit in other clinical platforms, including Tempus, when that context is part of the chart. Read those systems as a class of sources. Do not assume a vendor ranks pairs, invents a severity, or auto-holds a dose.

Run the load in this order:

1. Pull active MAR rows: drug, dose, route, start, and whether a dose is still scheduled.
2. Pull the new order with the same fields.
3. Normalize names to the identifiers the monograph uses, matching how your knowledge base keys pairs.
4. Request the monograph for that pair only.
5. If a monograph row returns, write the alert with both order cites and the monograph cite. If nothing returns, write empty.

Do not backfill a missing MAR row from a progress note or from memory. If the second drug is not on the MAR and not on the new order, there is no pair. An alert that names one drug and a class, "a statin," "macrolides," is the one-drug failure mode. Send it back until both order identifiers are present.

Empty is a valid result. It means the knowledge source did not return a matching pair for the two products you named. It does not mean safe. It does not mean override. It means you have nothing to cite.

Filling empty with a guessed interaction, a class-level warning you cannot attach to both orders, or a severity you inferred because the combination "feels high risk" fails the quality bar. The outcome is an alert that cites both drug orders and the interaction monograph, or it stays empty.

If the monograph is silent, stop. Do not promote a nearby pair, a different statin, a different macrolide, because it is in the same class. Do not copy a class warning onto this pair unless the monograph you loaded actually covers these two orders.

When the pair does match, keep the alert scoped to that pair. Do not bundle every other active drug into one undifferentiated interaction-risk banner. Additional pairs can be checked. They are separate alerts, each with two cites and a monograph, or each empty.

An interaction alert is a notice beside the order, not a stop. Pharmacy verification, nursing administration, and the prescriber's decision remain in place. If the workflow auto-cancels, auto-delays, or removes the dose from the MAR because a model raised a flag, therapy changed without a prescriber decision. Keep the scheduled dose until a human acts.

Inventing a severity is the same class of error. Monographs may already classify pairs. Carry the classification the source states, or carry none. Do not map "this looks major" onto a pair the source did not grade. Do not upgrade a stated moderate interaction because the patient is older. Age, renal function, and other charted factors can be listed next to the cite. They are not a license to rewrite the monograph's grade.

Related list errors look like safety too: citing a discontinued MAR row as if it were active; citing a PRN that was never given as if it were standing therapy; re-alerting on a pair already reviewed and documented this admission without restating that review. Fix the MAR load. Do not paper over a dirty list with a louder flag.

After a complete alert, the prescriber decides whether to change the new order, change the existing order, adjust monitoring, or proceed. Pharmacy can recommend. Nursing can withhold pending clarification when local policy says so. The model does not close the loop.

Write the alert so that decision is possible: both order cites, the monograph cite, and any patient-specific factors you actually loaded. Do not add a recommended action unless local protocol already maps that monograph to a standard pharmacy recommendation, and even then label it as protocol, not as model output.

pharmacogenomics-based drug selection can change which product is appropriate before this check runs. A genotype finding is not an interaction monograph. If the chart already has a pharmacogenomic result that affects one of the two drugs, include it as a patient-specific factor. Do not treat it as a second interaction.

adverse event signal detection looks backward for harm that already occurred. Real-time interaction monitoring looks forward at orders that have not yet been given. If an interaction was missed and a later signal appears, that is a different workflow. Do not merge the two outputs.

Keep the quality bar narrow on purpose. Load the MAR and the monograph. Alert only when both drugs and the monograph can be cited. Leave unmatched pairs empty. Let the prescriber decide.