Home

The adverse reaction rates MINA gave
matched the official product information exactly.

When you ask an AI for a drug's adverse reaction rates, you cannot always know which source the answer comes from. We saw how the source changes the answer in one example.

Two documents side by side: a promotional brochure with photos on the left and an official product information document with a table on the right. A line runs from the right-hand document down to a glowing brain icon.

We asked MINA for the adverse reaction rates of darolutamide, a drug used in prostate cancer. In its environment without internet access, MINA gave fatigue as 13.7%, and rash and hypertension in combination with docetaxel as 16.6% and 13.8%.

The official product information confirmed all three figures

We asked a general-purpose large language model with internet access to check the same answer. The general model said that two of the figures were wrong. It relied on the drug's promotional and information pages on the internet.

When the European Medicines Agency's official product information was checked, all three of MINA's figures were exactly right, and the general model acknowledged its error.

The source made the difference

The difference came from the source each model looked at. MINA's knowledge base carries the official product information of drugs. The model without internet access stayed with a more accurate source than the model with internet access.

Diagnosis in two difficult cases

The same general model also prepared two difficult cases and kept the answer key itself.

In the first case, in a diabetic patient whose kidney function was deteriorating rapidly, MINA recognised acute kidney injury, suspected metformin-associated lactic acidosis and infection together. It also caught the pitfall in the anticoagulant dose.

In the second case, in an elderly patient with yellow vision, nausea and a slow pulse, MINA placed digoxin toxicity first and linked it to amiodarone started five days earlier. It also saw that a regular, slow pulse in this patient with atrial fibrillation could point to a conduction block. Although the question did not ask for it, it calculated creatinine clearance and suggested adjusting the anticoagulant dose accordingly.

The physician makes the decision

MINA does not make clinical decisions. MINA brings information and its suggestion to the physician together with the source they are based on, and the physician always makes the decision.

To trust an answer in clinical decision support, you need to know which source that answer is based on.

You can read how MINA was built on the MINA page. If you would like to see it in your own institution, you can request a POC.

Related: TIS for physicians · MINA's clinical knowledge measurement

Source: The rates appear in section 4.8, undesirable effects, of the official product information (summary of product characteristics) published by the European Medicines Agency for darolutamide. This section gives 13.7% for fatigue and asthenic conditions, and 16.6% for rash and 13.8% for hypertension in combination with docetaxel. The comparison with the general model and the two cases come from a test conversation held in November 2026. The general model prepared the cases and their answer keys.

SSS

What were the rates MINA gave compared with?

The rates were compared with section 4.8 of the European Medicines Agency's official product information. The rates MINA gave for fatigue (13.7%), and for rash (16.6%) and hypertension (13.8%) in combination with docetaxel, are identical to the values in that document.

Why did the general model reach a different conclusion?

The general model relied on the drug's promotional and information pages on the internet. MINA's knowledge base carries the official product information of drugs.

Does MINA make clinical decisions?

No. MINA brings information and its suggestion to the physician together with the source they are based on. The physician always makes the decision.