This medicine is called Tacrolimus

It is a medicine taken every day by people who have received an organ transplant (kidney, liver, and so on).

Our body mistakes the newly transplanted organ for a strange “intruder” and tries to attack it. Tacrolimus calms that attack so the new organ can settle in well.

The right amount is what matters

If there is too much, the kidneys can be harmed or side effects appear;
if there is too little, the body rejects the new organ.

So the drug level in the blood must be kept right inside a narrow safe zone.

But it differs from person to person

Even with the exact same dose, the blood level is different for every person.

Age, weight, genes, other medicines taken together… it varies for countless reasons. So predicting “what will this patient’s level be next time?” in advance is very hard.

The method so far — the average calculator (PopPK)

The traditional method (population pharmacokinetics, PopPK) calculates the level based on an “average patient.”

Upside: it can be used right away, even for a brand-new patient.
Downside: it cannot capture that person’s individual differences well.

Our idea ① — start from the average

A brand-new patient has no data of their own yet. So we cannot build a personalized prediction.

So we took the average prediction (DVI), which can be computed for anyone, as a solid starting point.

Our idea ② — learning from the last error

From the second visit on, we get a clue: “how far off was the prediction last time?”

We look at this last error (ERR_LAG1) and gently nudge the next prediction. As visits pile up, it gets closer to a personal fit.

Three AI friends lend a hand

Three artificial-intelligence models read the patient’s past visit records and predict the next level.

All of them use our idea — “average prediction + last error” — as their ingredients.

Result — combining the two fits better!

We compared how far the predictions were off from reality (the error — lower is better).

When we combined the average prediction (PopPK) with AI, all three models had less error than PopPK alone. Among them, LightGBM was the most accurate. 🏆

* Numbers are the prediction error (RMSE, ng/mL) — the lower the bar, the more accurate the prediction.

The most important clue? The last error!

When we opened up what the AI looked at to decide (a SHAP analysis), all three models picked “the last error” as the most important clue.

Remarkably, that error was on average about a month old — and it still mattered. This suggests a person’s “habit” can be relatively stable between visits.

So, what is good about it?

✅ It can be applied right away, even to a brand-new patient.

✅ As visit records pile up, it gets closer to a personal fit.

✅ It works without any complex per-patient re-computation.