A warfarin dose, and how far to trust it.

WarfaRisk predicts a weekly warfarin dose and a 90% range, then checks how well that range held up for the patient's ancestry group.

Supported by

  • Alfaisal AI Center
  • Alfaisal University
Example patient
Weekly dose
31.8 mg
90% range
13.0 to 50.6 mg

Held upFor Han Chinese patients the true dose was inside the range 100% of the time. Goal: 90%.

Tablets are an illustration. Research prototype.

A good score can come from a small test.

Warfarin is hard to dose. The right dose can differ ten times between patients, and too much or too little is dangerous.

Best published error
0.77 mg/week
Tested on only 64 patients.
Our best model
8.55 mg/week
Tested on 1,207 patients. A careful test is less flattering.
64 patients, the small test1,207 patients, our test

So we built a careful test, in nine steps.

Steps 1 to 6 are the evaluated core. Steps 7 to 9 are prototype extensions.

  1. 1

    Clean the data

    Fit every fix on training patients only, then run 11 checks for leaks.

  2. 2

    Set the bars to beat

    Three simple methods, including the published IWPC formula.

  3. 3

    Compare nine models

    Each one with clinical details, genes, and both together.

  4. 4

    Leave one group out

    Train without one ancestry group, then test on it.

  5. 5

    Check the intervals

    Check how often the real dose lands inside, by group.

  6. 6

    Explain each dose

    Show which inputs pushed the dose up or down.

  7. 7

    Follow doses over time

    Test on hospital records with repeated INR checks.

  8. 8

    Report fairness one way

    Same tables for every group, in every result.

  9. 9

    Write the reason

    A short plain explanation that cites its sources.

The first result: genes help, but not enough yet.

On 1,207 test patients, every model that used genes beat the same model without them.

Average error

8.55mg/week

Lower is better. Best model, public IWPC data.

Doses within 20% of the true dose

46.6%

The usual bar is 50%. We did not reach it.

Dose variation explained

47.5%

Genes added roughly 10 to 20 points over clinical details alone.

Average error by method, mg/week

  • Guess the median dose12.34
  • Clinical details only, linear model10.86
  • Published IWPC dosing formula9.18
  • Best WarfaRisk model8.55

11 leak checks before any model trains

Training and test patients never overlap.

An average says nothing about who the tool works for. So we checked.

The tool says how sure it is. For some groups it is too sure.

Each dose comes with a range that should hold the true dose 9 times out of 10. Across all patients it does: 89.8%. A common shortcut, averaging five neural networks, reached only 24.9%.

Share of true doses inside the 90% range, by ancestry group, highest first. The target is 90%.
GroupPatientsInside the range95% intervalResult
All patients120789.8%not reportedMet the target
Korean5396.2%87 to 99.5Met the target
Japanese17095.9%91.7 to 98.3Met the target
White49391.3%88.4 to 93.6Met the target
Chinese2889.3%71.8 to 97.7Below target, could be chance
Caucasian12486.3%79 to 91.8Below target, could be chance
Intermediate1485.7%57.2 to 98.2Below target, could be chance
Black7382.2%71.5 to 90.2Below target, could be chance
Black or African American5269.2%54.9 to 81.3Below target, not chance
African-American2462.5%40.6 to 81.2Below target, not chance
Dot: share of true doses inside the range. Line: the 95% interval, longer for smaller groups. Met the target Below it, could be chance Below it, not chance

Split by ancestry, it does not

For the two red groups, about 1 true dose in 3 fell outside the range, against about 1 in 10 overall. Every range was the same width, about 37.7 mg/week, so the tool sounded just as sure.

Calibrating each group helps

  • Black82.2% 93.2%
  • African-American62.5% 87.5%

Ranges get wider, and groups too small to measure stay out of reach.

Why a dose comes out as it does, and where the model runs short.

The tool shows which inputs pushed each dose up or down. Across patients genes carry about 64% of that weight, led by VKORC1.

  • VKORC1 (AA)Gene8.1
  • VKORC1 (AG)Gene3.9
  • AgeClinical3.3
  • WeightClinical2.0
  • HeightClinical1.6
  • CYP2C9 (*3 type)Gene1.1
  • CYP2C9 (*2 type)Gene0.9

Average push on the dose. Top 7 of 11 inputs, 1,780-patient cohort.

Here is the tool, with four real examples.

Pick a patient and follow the three screens a clinician would see. Dose, range and input effects come from the real model; group figures come from the papers. Research prototype, not for clinical use.

WarfaRisk
Research prototype

Example values, so they cannot be edited. Pick another patient above to see a different result.

Patient

Age
68years
Weight
72kg
Height
168cm
Target INR
2.5range 2.0 to 3.0
Reason for warfarin
Atrial fibrillation
Ancestry
Han Chinese

Other medicines

Amiodarone
No
Statin
No
Enzyme inducer
No

Genes

CYP2C9
*1/*1
VKORC1 -1639G>A
AG

Optional. Without genes the model uses clinical details alone.

Step 1 of 3

Common questions.

Short answers, with the numbers from the papers.

WarfaRisk is an open research project that predicts a weekly warfarin dose with a 90% range around it, then checks how well that range held up for each ancestry group. It is described in two peer-reviewed papers from 2026 and shown here as a prototype. It is not for clinical use.

No. WarfaRisk is a research prototype, trained on public research data and not tested in patients. The dose and range shown on this page must not be used to make dosing decisions. Dosing decisions belong to a clinician.

On 1,207 test patients, the best model's average error was 8.55 mg per week, against 9.18 for the published IWPC dosing formula. 46.6% of doses were within 20% of the true dose, below the usual bar of 50%. These figures are from the Journal of Personalized Medicine paper.

No. The 90% range held for 89.8% of all patients, but for patients labelled Black or African American it held for 69.2% (52 patients), and for African-American patients 62.5% (24 patients). Both are below the 90% goal. Calibrating each group on its own raised the Black label from 82.2% to 93.2%, and Black or African American from 69.2% to 78.8%.

WarfaRisk is a research project from Alfaisal University and partner institutions in Saudi Arabia. The project leads are Dr. Mariam AlEissa, Dr. Areej Al-Wabil and Dr. Abdullah N. Alkattan. The team page lists everyone who worked on it, with links to their profiles.

Yes. The pipeline code, the prototype app, the interface demo and the trained models are open. Links to the papers, the code and the models are in the Papers section of this page.

Everything above is in two open papers.

Both are free to read. The code and trained models are open too.

Journal of Personalized Medicine, 2026, 16(10), 531

Beyond Mean Absolute Error: Ancestry-Stratified Calibration and Explainability for Warfarin Dosing Models

Accuracy, who the tool misses, how sure it is, and why, on the same patients.

Read the paper

DOI 10.3390/jpm16100531

AI, 2026, 7(10), 431

WarfaRisk: A Reproducible, Fairness-Aware Machine Learning Pipeline for Warfarin Dose Prediction (Prototype)

The tool: the nine steps, the three screens, and what it takes to reach a clinician.

Read the paper

DOI 10.3390/ai7100431

Supported by

The authors thank the Alfaisal AI Center at Alfaisal University for infrastructure support.

An average can hide who a tool fails. Check every group.