An anticoagulated patient wants a "natural" remedy: the interaction AI explains backwards

Reading time: about 11 minutes

The situation. A man of about seventy, a regular customer who has been on acenocoumarol (a vitamin K antagonist, like warfarin) for years, comes in with a box of a supplement "for mood" recommended by a neighbour — it contains St John’s wort — and also asks about some omega-3 capsules "for the heart". He wants to know whether he can take both.

This is the classic interaction query, and here AI has a very particular failure: it usually gets right that there is an interaction and gets the direction wrong. Because the patient is anticoagulated, the model links "risk" with "bleeding" and applies it to everything. With St John’s wort it is the other way round: the problem is that the anticoagulant stops working as well. A reply pointing the wrong way does not just fail to help: it has you watching for the wrong thing.

What you need to hand: The supplement interaction checker on this site — it works from a curated dataset and says where each piece of information comes from — the box itself so you can read the ingredients, and the lesson on AI and the interaction checker, which explains why the dataset comes first. Nothing identifying the patient goes into the chat: for an interaction, the active ingredients are enough.

Step by step

Read the ingredients, not the product name

A supplement "for mood" can contain almost anything, and the brand name does not tell you what. The first step is to turn the box over and read the full list of ingredients: here it shows St John’s wort (Hypericum perforatum) and a vitamin or two. The omega-3 capsules state their daily dose of EPA and DHA on the label.

Note the daily omega-3 dose he would get by taking it as the pack says. This is not a detail: for omega-3 with anticoagulants, the dose is what changes the answer, as you will see in the next step.

And ask what else he takes. Someone on acenocoumarol often takes several other things, and St John’s wort has important interactions with other medicines besides the anticoagulant.
Run it through the reviewed dataset first

Before asking any chat, enter acenocoumarol, St John’s wort and omega-3 into the interaction checker. What it returns for these pairs comes from the curated dataset:

Acenocoumarol + St John’s wort: Major. Mechanism: St John’s wort is an enzyme inducer and reduces the effect of the anticoagulant, with a risk of thrombosis. Action: stop the St John’s wort and watch the INR.

Acenocoumarol + omega-3: Moderate. At high doses — above about 3 g a day — it adds an antiplatelet effect of its own. At usual supplement doses no adjustment is needed; at high doses, watch the INR and minor bleeding.

Notice that the two interactions point in opposite directions: one lowers the anticoagulant’s effect and the other, at high doses, adds bleeding risk. That is exactly what a generic reply mixes up.
Use AI to understand the mechanism, giving it the data

With the dataset’s information in front of you, AI is useful for what it does well: explaining the mechanism calmly, or helping you explain it to the patient. But you give it the information; you do not ask for it.
```
Explain to me in 6 lines, for my own understanding as a pharmacist,
why St John’s wort REDUCES the effect of a vitamin K antagonist
anticoagulant such as acenocoumarol.

Information already checked in a reviewed source:
- Interaction: Major.
- Mechanism: enzyme inducer → less anticoagulant effect →
  risk of thrombosis.
- Action: stop the St John’s wort and watch the INR.

Do not change the direction of the interaction. Do not add any
recommendation about the anticoagulant dose.
```
The instruction not to change the direction looks unnecessary and is not: it is exactly the error you will see in the example below. And the one about not touching the anticoagulant dose protects against the other typical error: suggesting an adjustment that only whoever manages his treatment can decide.
Decide what you tell the patient, and what you do not

With a Major interaction, the conclusion for St John’s wort is clear: you do not recommend it. And if he has already started it, telling him to stop is not enough: when an enzyme inducer is withdrawn, the anticoagulant’s effect can rise again, and that too is monitored through the INR. It is a reason for whoever manages his anticoagulation — his doctor or his anticoagulation clinic — to know, not for anything to be adjusted at the counter.

With omega-3, it depends on the dose. At the usual dose of a supplement, the dataset says no adjustment is needed; even so, it is worth it being recorded at his clinic and him knowing what to watch for (bruising, bleeding gums). At high doses, the advice is to watch the INR, and that is again a matter for his clinic.

And there is the underlying question: why does he want something for his mood? If he is feeling low in a way that worries him, that deserves a conversation and probably an appointment with his doctor, more than a herbal product.
Prepare the explanation for him in his own words

Here AI genuinely saves you time: a simple explanation, without jargon, of what you have decided.
```
Write 5 lines explaining to an older man who takes an anticoagulant
why he should NOT take a supplement containing St John’s wort: it makes
his medicine work less well, and that can be dangerous. Tell him to
mention it to his doctor or his anticoagulation clinic. Warm, calm tone.
Do not mention doses.
```
Read the result before using it. If it says "may increase the risk of bleeding" about St John’s wort, it is wrong and needs correcting: it is the phrase the model reaches for most readily whenever it hears "anticoagulant".

A full example, and what is wrong with it

This is what a model replied to "a patient on acenocoumarol wants to take St John’s wort and omega-3 — is there an interaction?", given no prior information. It is shortened and not polished.

What it replies

"Yes, both combinations have relevant interactions with acenocoumarol.

St John’s wort enhances the anticoagulant effect and increases the risk of bleeding, so it should be avoided or, if used, the acenocoumarol dose should be reduced and signs of bleeding monitored.

Omega-3 fatty acids are contraindicated in anticoagulated patients because of their antiplatelet effect, so their use is not recommended."

And now, what is wrong with that reply:

The direction of the interaction is backwards. St John’s wort does not enhance the anticoagulant: it reduces it, through enzyme induction, and the risk is thrombosis. The reply has applied the "anticoagulant + anything = bleeding" template to a case that goes the other way. You would be watching for bruises while the real risk is a clot.
It suggests reducing the anticoagulant dose. Besides being the wrong direction — if St John’s wort reduces the effect, lowering the dose makes it worse — it is a decision only whoever manages his treatment, with the INR in front of them, can make. A suggested adjustment in a chat reply is exactly what must not leave the counter.
It overstates omega-3. "Contraindicated" is not what the dataset says: the interaction is Moderate, relevant at high doses, and at usual supplement doses no adjustment is needed. Turning "watch at high doses" into "contraindicated" is also an error, just in the cautious direction, and it makes the patient distrust everything else when he reads something different.
It does not say where anything comes from. No dataset, no SmPC, no source. With a reply like that there is no way of knowing which of the three claims is wrong except by checking them all, which is what you had to do anyway.

A reply that is right in general — "there are relevant interactions" — and wrong on every specific: the direction, the action and the severity. And the correct opening is what makes people believe the rest.

That is why the case runs in this order: first the dataset, which gives the direction and severity with its source; then AI, to explain what you already know. The other way round, AI tells you what to look for and what you look for is wrong.

And if your pharmacy is not like that

If the patient takes a direct oral anticoagulant instead of acenocoumarol. St John’s wort also interacts with several of them, through the same induction mechanism, but monitoring is not done with the INR, so the follow-up is different. Check it in the checker with the specific active ingredient: do not assume the acenocoumarol answer applies unchanged.

If he also takes an antidepressant. The dataset lists a Major interaction between St John’s wort and SSRIs, with the action of stopping the St John’s wort. That is a second reason, independent of the anticoagulant, and a common one: many people look for something natural "for mood" precisely because they already take an antidepressant and feel it is not enough.

If the person asking is a woman taking the oral contraceptive pill. St John’s wort also has a Major interaction with it in the dataset: it can reduce its effectiveness. It is a question almost nobody asks at the counter and one worth asking whenever someone wants St John’s wort.

When it does not work first time

The checker does not find the product name.

That is normal: it searches active ingredients, not supplement brands. Read the ingredients and search for "St John’s wort" or "hypericum". If the product does not state the amount of each ingredient, that too is information to give the patient.

The AI insists St John’s wort increases bleeding even when I give it the data.

Do not argue: open a new conversation and paste the dataset’s information at the start, before the question, with the instruction not to change the direction. If it still changes it, do not use that reply for anything: write the explanation yourself with the mechanism from the dataset.

The patient says he has been taking the St John’s wort for a month.

Then what matters is not only that he stops it, but that his anticoagulation clinic knows, because his INR may have dropped over these weeks and may rise when he stops. Note it down, explain it to him and refer him to his doctor or clinic so they can decide when to check the INR.

I do not know what dose of omega-3 he is taking.

Look at the label: the daily dose of EPA and DHA is usually given. If it is a normal supplement dose, the dataset says no adjustment is needed; if he is taking several capsules more than directed, that is a different conversation.

Before you use what it gave you

You have read the product’s full list of ingredients, not just its name.
You ran the active ingredients through the checker before asking the AI.
You know the direction of each interaction: St John’s wort reduces the anticoagulant’s effect; omega-3 at high doses adds an antiplatelet effect.
You have not suggested any change to the anticoagulant dose.
If he was already taking it, you referred him to his doctor or clinic to decide about the INR.

Everything is bleeding. Faced with an anticoagulant, the model tends to attribute any interaction to bleeding risk. Some go the other way, and those are exactly the ones it matters most not to confuse.

It suggests adjusting the anticoagulant dose. It is advice that sounds technical and cannot be given without the INR and without managing the patient’s treatment. No chat reply should contain it, and none should leave the counter.

Severities overstated or understated. "Contraindicated" where it is moderate, "no problem" where it is major. Without the source you cannot tell which it has done, and in the same reply it can do both.

Now, with your own. Open the interaction checker and enter three combinations with an anticoagulant that you see often in your pharmacy. Read the direction and the action for each one. Then ask a chat about the same combinations, giving it no data, and compare. Note down the ones it got backwards: those are the ones worth having clear before a patient asks.

← All practical cases Back to the school →

An anticoagulated patient wants a "natural" remedy: the interaction AI explains backwards

Step by step

Read the ingredients, not the product name

Run it through the reviewed dataset first

Use AI to understand the mechanism, giving it the data

Decide what you tell the patient, and what you do not

Prepare the explanation for him in his own words

A full example, and what is wrong with it

And if your pharmacy is not like that

When it does not work first time

Before you use what it gave you