The Polypharmacy Trap in the Geriatric Ward

A 78-year-old man, four years on warfarin for atrial fibrillation, admitted with a surgical site infection. The on-call resident is about to add metronidazole. It is the obvious empirical choice. He is tired enough that he might prescribe it without thinking.

He types the order into the system. Below the new order line, a single sentence:

Possible interaction: metronidazole + warfarin. INR may rise within 3–5 days. Consider azithromycin if anaerobic coverage is not essential. Otherwise: drop warfarin to ½ dose, check INR daily, hold warfarin if INR > 4.

He switches to azithromycin. Three days later the patient's wound is clean and his INR is 2.4 — unchanged.

The interaction was not obscure

Warfarin and metronidazole is not a subtle finding at the edge of the literature. It is in every formulary, it is taught, and the resident in this case could have described the mechanism if asked in a viva that morning.

That is the entire point, and it is worth sitting with rather than moving past. The problem is not that the knowledge is missing. It is that the knowledge has to be retrieved at the moment of writing the order, by a person who is at hour fourteen and is thinking about the wound, the fever, and the four other patients he has not seen yet. Recall is not free, and it fails silently. Nothing in the resident's experience of writing "metronidazole" flags that a retrieval has just failed to happen.

Any intervention that assumes the fix is education is aimed at a problem that does not exist.

The combinatorics

The average elderly patient on a medical ward in India is on 6–8 chronic medications. The average admission adds 2–4 acute ones. The combinatorics of interactions is large enough that no resident can hold the full picture in their head.

Put numbers on it. Eleven concurrent drugs is 55 distinct pairs. Check each pair for a clinically meaningful interaction and you have performed 55 lookups for one patient, on a ward with twenty-eight of them. Nobody does this. Nobody has ever done this. What actually happens is that the doctor checks the two or three pairs that occur to them, and the pairs that occur to them are the famous ones — which means the catch rate is highest exactly where the risk is best known and lowest everywhere else.

And the pairs are not static. Every new admission drug re-opens all of them. The patient who was safely reviewed on admission is a different combinatorial problem by day three.

The published cost

The published data on this is grim. Hajjar and colleagues (Am J Geriatr Pharmacother, 2007) found that adverse drug events were responsible for 16% of hospital admissions in patients over 65 in their cohort. Of those, an estimated 60% were potentially preventable — the interactions were known, the at-risk patients were identifiable, the system simply did not catch them at the point of prescribing.

"Known, identifiable, uncaught" is the phrase to take from that. This is not a research frontier. It is a delivery failure.

Why the catch has to happen at the keyboard

Drug-interaction checking is a feature in most pharmacy software. It is rarely a feature embedded in the act of prescribing on the ward. The resident on the ward is using paper, an EHR for documentation, and a separate drug reference if they bother. The interaction check happens, if at all, when the pharmacist dispenses — which is too late to influence the prescribing decision without a friction-laden callback.

Trace what that callback actually costs. The pharmacist notices at 3 p.m. and calls the ward. The resident who wrote the order is in theatre, or off shift. Someone takes a message. When it reaches the prescriber, they have to reload the entire clinical context — why this antibiotic, what the alternative would cost in coverage, whether the anaerobes actually matter here — a context that was fully loaded and free at the moment they wrote the order, and now has to be rebuilt from the chart.

The catch is not wrong. It is just expensive, and expensive catches get deferred, and deferred catches sometimes do not happen.

Why it must not be a modal

The obvious engineering response is a blocking dialog, and it is the wrong one.

Interaction alerts in prescribing systems are overridden at rates high enough, in the published literature, that the alert has become close to decorative — clinicians learn the dismissal keystroke and execute it before reading, because the great majority of what fires is irrelevant to the patient in front of them. A blocking alert that is dismissed reflexively is worse than no alert, because it consumes the attention budget that a real warning would have needed and it lets everyone believe a safety system is in place.

In Medows, the AI consult is grounded in the patient's active medication list. When the resident considers an addition, the interaction check runs automatically — a single sentence, in line with the order, citing the relevant evidence. It is not a modal. It is not a blocking dialog. It is the equivalent of a pharmacist quietly noting something next to the order while the resident is writing it.

What a useful interaction line contains

Look again at the sentence that changed the outcome, because its structure is not accidental. It carries four things, in this order:

1. What interacts, and what it does — the INR may rise. Not a severity grade, not a colour code: the actual consequence.
2. On what timescale — 3 to 5 days. This is the part almost every interaction database omits, and it is the part that determines whether you can safely proceed and monitor.
3. A concrete alternative — azithromycin, with the clinical condition under which the swap is reasonable.
4. A plan for proceeding anyway — half dose, daily INR, hold above 4. Because sometimes the anaerobic cover genuinely is essential, and an alert that only says "don't" is useless to the doctor who has to.

A warning that stops at step one is asking the resident to do the thinking that the tool was supposed to save. The point is not to replace the pharmacist. The point is to have the catch happen during the prescribing decision, when the resident still has the full clinical context fresh.