AI School · Level 3 · Lesson 2

Medication review

Léelo en español →

A medication review is not "run the list through an interaction checker". It is a six-step procedure, and AI genuinely helps with two, gets in the way of one and is no use at all in the other three. Knowing which is which is the whole lesson.

And the hard step is not the analysis: it is the list. What the person actually takes is almost never what the prescription record says. Missing from it: what they self-medicate, what a relative gives them, the supplement bought somewhere else, and what they quietly stopped two months ago. With the list wrong, everything that follows is precise and false.

The six steps, and where AI comes in

  1. Gather what they actually take

    The prescription record + what is in the bag + what they tell you. All three, and all three give different lists.

    ✗ AI does not help: this is asking well, and it is the next lesson.

  2. Structure the list

    Active ingredient, dose, schedule, what for, since when and who prescribed it. Turning "a white tablet in the morning" into a row with its six fields is mechanical, tedious work.

    ✓ Here AI genuinely saves time — it is reformatting, not deciding.

  3. Look for duplication

    Two drugs from the same therapeutic class. It is the most frequent finding of a review and the most invisible: both are correctly prescribed, the problem is that both are there. Usually one was meant to stop when the other started.

    ✓ AI groups by class well, and that is exactly what is hard to see by hand.

  4. Cross-check interactions

    Every drug against every other. With 9 drugs that is 36 pairs; with 12, it is 66. It cannot be done by hand, which is why this step gets skipped so often.

    ✗ Do not ask a general-purpose chat: use a tool that crosses the pairs and tells you which ones it could NOT check.

  5. Appropriateness for this person

    Age, renal function, swallowing, who prepares their medication. A drug that is right for the indication may not be right for the person taking it.

    ✗ AI answers about the average case, and here what matters is what makes them not the average case.

  6. Write the note to the prescriber

    Short, with the findings ordered by severity and one concrete proposal each. The decision is the doctor's; the note is yours.

    ✓ AI drafts this well — with the findings you give it.
A pair that has not been looked at is not a clean pair. If the tool could not check three of the thirty-six pairs, those three do not go in with the checked ones: they go separately and they get said. "No relevant interactions" on its own, with pairs unchecked, closes the review falsely — and a falsely closed review is worse than none, because nobody opens it again.

What the pharmacy does NOT decide

Finding a duplication is not withdrawing a drug. What comes out of a review is a proposal to the prescriber, and there is a clinical reason as well as a professional one: stopping a long-standing benzodiazepine, or an antidepressant, abruptly carries its own risk. The sentence to write is "review with the prescriber", not "stop".

Save it, and save separately what you did not check

The same review gets picked up again: the patient comes back, the doctor calls, somebody has to explain why something changed. Being able to retrieve it is half its value — and it is free with any account.

And what was left unchecked is stored SEPARATELY from what was clean. They mean opposite things, and stored it is worse than on screen: a review reopened three months later would claim to have checked something that was never checked.

A full example, and what is wrong with it

The lesson says step 4 — cross-checking interactions — is not a question for a general chat. This is what happens when you ask it anyway, on a synthetic nine-drug list, which is 36 pairs.

The list pasted in: enalapril · furosemide · metformin · omeprazole · sertraline · fluoxetine · lorazepam · diazepam · ibuprofen as required.

What it replies

I have reviewed all 36 possible combinations between the nine drugs. The relevant findings are:

1. Ibuprofen + enalapril + furosemide: triple whammy, risk of renal deterioration. I recommend stopping the ibuprofen and replacing it with paracetamol.

2. Omeprazole + sertraline: possible increase in sertraline levels through CYP2C19 inhibition. Monitor.

The remaining combinations show no clinically relevant interactions. The regimen as a whole is appropriate.

And now, what is wrong with that reply:

  1. It has missed two duplications that are in plain sight. Sertraline and fluoxetine are two SSRIs. Lorazepam and diazepam are two benzodiazepines. That is not an interaction between different drugs: it is the same class twice, which is the commonest finding in a review and the one the lesson puts first. And notice why it slips through: both are correctly prescribed, neither looks odd on its own, and an interaction checker that only looks at pairs of different active ingredients does not show it.
  2. "I have reviewed all 36 combinations" is something it cannot know. There is no way for it to tell you which ones it looked at. No list, no trace, no way to ask for pair number 17. It has written a sentence that fits the question — counting pairs is arithmetic and arithmetic it can do — and that reads like a guarantee of coverage. That sentence is exactly why the lesson insists on a tool that crosses the pairs and says which ones it could not check.
  3. "The rest show no interactions" is the opposite of the truth. The correct wording would be "I found nothing in the rest", which is a different statement. What is happening here is the review being closed: whoever reads that line in three months will understand that the remaining 34 pairs were checked and came back clean. A pair that has not been looked at is not a clean pair, and mixing them is what turns an incomplete review into one closed under false pretences — which is worse than none, because nobody reopens it.
  4. And "stop the ibuprofen" is not the pharmacy's to decide. The finding is good — that combination is textbook — and the conclusion oversteps by one word. What comes out of a review is a proposal to the prescriber, and the wording is "review with the prescriber". This is not a formality about scope: written as it is, it ends up said at the counter verbatim, and the person hearing it stops taking something without anybody having assessed why they were taking it.

What is worth keeping is how the work divides. That same list, run through this site's interaction checker, comes back pair by pair, severity first and — crucially — separating "checked and clear" from "not checked". The class duplications show up because the dataset has rules written for exactly that.

And then the chat becomes useful again, at step 6: give it the findings yourself and ask for the note to the prescriber, ordered by severity with one proposal per finding. There it writes well and there is nothing it can invent, because you supplied the findings.

And if your pharmacy is not like that

If the review is for a care-home resident. The list arrives already written, which looks like an advantage and is the trap: that list is already somebody's interpretation, and what is missing — what the family gives at weekends, the syrup the daughter bought — is on no sheet. Step 1 is still the hard one, only now you ask the staff who administer rather than the patient. And beware the shortcut of uploading the whole sheet to a chat: that is twenty identified people at once.
If there are only four or five drugs. With five it is ten pairs and you can do it by hand, so AI adds less at step 4 and just as much at steps 2 and 6. But duplication is still the star finding and with few drugs it is seen less, not more: with nine you look for patterns, with five you assume they are fine because there are so few. The run-through by therapeutic class takes thirty seconds and that is where the result is.
If the patient takes supplements and herbal products. This is half the real problem and it almost never makes the list, because it is not considered "medication". Ask explicitly and by category — "anything for sleep? anything for your mood? anything from the health shop?" — which is what the next lesson teaches. And once on the list, treat them as drugs: St John's wort has interactions that change a regimen, and it is on no prescription record.
If you do reviews inside a funded service with a record. Then what gets written down matters as much as what you found, and two things always go in: the source of each finding and what was left unchecked. "Interaction X per the product information" and "3 pairs not checked" are the two lines that make that review worth something six months later. What does not go in is which tool you used — it informs nobody reading the sheet afterwards and it changes how everything else on it is read.

When it does not work first time

The patient cannot remember what they take and has brought nothing.
Then there is no review that day, and saying so is part of the job: a review built on an incomplete list is precise and false, and it goes on record as done. What you can do is set up the next one — the prescription record gives you half, and for the other half you book them with "bring the bag with everything in it, including anything from the health shop and anything you have stopped". That sentence is what produces the good list.
I find a duplication and the doctor does not reply.
It is still better than not finding it, and two things lift the reply rate a great deal: a note that is short and carries a concrete proposal — not "there is a duplication" but "two SSRIs have coexisted since March; should one be withdrawn?" — and saying what you have already checked. If there is still no reply, record that it was communicated and when. That is not bureaucracy: it is what turns an attempt into a documented finding.
The tool leaves pairs unchecked and I do not know what to do with them.
First, be glad it tells you — that is exactly what a chat does not do. Then: unchecked pairs get looked up by hand in the product information if they involve anything risky (anticoagulants, antiarrhythmics, lithium, immunosuppressants), and get recorded as pending if not. What you never do is add them to the clean ones. If there are a lot, it is usually a connection or quota problem: run it again later rather than calling the review closed.
What if the review finds nothing?
It happens often and it is a result, not a failure — as long as you can say what was looked at. "36 pairs reviewed, no findings, 0 unchecked" is useful and defensible; a bare "no relevant interactions" says nothing and sounds like nothing was looked at. And even with no findings, two valuable things remain: the real list of what they take, which almost never existed before, and the conversation you had to get it.

Before you call this learned

← Previous: reading a medical report Next: preparing the conversation →