warfarisk-autogluon-6256

Beyond MAE: Phase 3 AutoGluon (IWPC-6256)

<p align="center"> <img src="https://placehold.co/800x200/dbeafe/1e40af?text=Beyond+MAE+%E2%80%94+AutoGluon+IWPC-6256" alt="Beyond MAE: Phase 3 AutoGluon IWPC-6256"> </p>

This repository contains the Phase 3 winning model from the pipeline described in "Beyond Mean Absolute Error: Ancestry-Stratified Calibration and Explainability for Warfarin Dosing Models", a nine-phase reproducible ML pipeline for warfarin dose prediction, trained on the IWPC-6256 cohort. An AutoGluon extreme_quality predictor on the combined clinical+genetic feature set reaches MAE 8.552, R² 0.475, PW20 0.466, the best of 9 architectures × 3 feature sets evaluated on this cohort, narrowly ahead of TabPFN (0.463) and CatBoost (0.454).

By: Hassan Barmandah, Omar Abdullah Bawazir, Siraj Aldeen Marghalani, Moath Shaat, Abdullah N. Alkattan, and Mariam M. AlEissa (corresponding author), Alfaisal University, Riyadh (with Umm Al-Qura University, Saudi Electronic University, Ministry of Health, Public Health Authority, King Khaled Eye Specialist Hospital Research Center, King Abdullah Petroleum Studies and Research Center)

This is a research artifact, not a validated clinical tool. It has not been evaluated prospectively and has no regulatory status.

---

Model Description

This model is an AutoGluon TabularPredictor fit with the extreme_quality preset on IWPC-6256 (n=6,037 after cleaning, n_test=1,207, fixed 80/20 patient-level split, seed 20260725), using the combined feature set (clinical + CYP2C9/VKORC1 genotype consensus). It is the best-performing model of 9 architectures evaluated across 3 feature sets on this cohort in the Phase 3 nested ablation.

Key Contributions

• Best Single Model on IWPC-6256: MAE 8.552 / R² 0.475 / PW20 0.466, ahead of TabPFN, CatBoost, and 6 other compared architectures
• Combined-Feature Advantage: beats the same architecture family's clinical-only and genetic-only results, confirming pharmacogenomic features add real signal beyond clinical variables alone
• Automated Ensemble, Documented Configuration: extreme_quality internally builds a weighted ensemble across its own model portfolio (final WeightedEnsemble_L2 used RealTabPFN-v2_c1 for this cohort); see Training Procedure for the exact time budget and excluded model types used to produce this result

---

🚀 How to Use

from autogluon.tabular import TabularPredictor
from huggingface_hub import snapshot_download

local_dir = snapshot_download("HassanB4/warfarisk-autogluon-6256")
predictor = TabularPredictor.load(local_dir)
prediction = predictor.predict(your_dataframe)  # combined clinical+genetic features, see Training Data below

Input features: age (band), weight (kg), height (cm), indication for anticoagulation, comorbidity flags (diabetes, heart failure, smoking), comedication flags (amiodarone, statins, enzyme inducers), CYP2C9 star-allele consensus, VKORC1 -1639 (rs9923231) consensus. See HasanBGit/WarfaRisk's src/warfarisk/phase3_genetics_ablation.py for the exact feature-engineering pipeline.

---

⚙️ Training Procedure

Training Data

Hyperparameters

TABDPT/TABICL/TABM/MITRA were excluded after a first attempt: TABDPT alone overran its allotted time slice by over 15 minutes with no way to interrupt mid-fit (AutoGluon's time_limit is checked only between models); re-run clean after exclusion. Full run notes in docs/RESULTS.md and docs/KNOWN_ISSUES.md.

Models Compared (combined feature set, IWPC-6256)

---

📊 Evaluation Results

PW20 = proportion of predictions within 20% of the true dose (IWPC's own clinical-acceptability threshold is ≥0.50: this model, the best in the sweep, comes close but does not yet clear it).

Comparison with Phase 1 Baselines (IWPC-6256)

---

⚠️ Limitations

• Research-Only: not validated prospectively, no regulatory status
• PW20 Below Clinical Threshold: 0.466 vs. the IWPC-established ≥0.50 acceptability threshold
• Fragmented Ancestry Labels: IWPC-6256's Black/African-American patients are split across three overlapping labels; see docs/RESULTS.md for this project's Phase 4 ancestry-holdout and Phase 5 per-subgroup calibration findings, which this specific model was not itself re-evaluated under (Phase 4/5 used a stacking-ensemble/CatBoost substitute; see docs/RESULTS.md for exactly which model was used for which analysis)
• No Raw Data Shipped: the code repository does not redistribute IWPC data; see data/DATA.md

---

🙏 Acknowledgements

We thank the PharmGKB / International Warfarin Pharmacogenetics Consortium for the IWPC dataset, and the AutoGluon team for the AutoML framework this model was built with.

Related Links

• AutoGluon
• PharmGKB
• Code repository

---

📜 Citation

This model is described in the following article, published in the Journal of Personalized Medicine.

Barmandah, H.; Bawazir, O.A.; Marghalani, S.A.; Shaat, M.; Alkattan, A.N.; AlEissa, M.M. Beyond Mean Absolute Error: Ancestry-Stratified Calibration and Explainability for Warfarin Dosing Models. J. Pers. Med. 2026, 16 (10), 531. https://doi.org/10.3390/jpm16100531

@Article{jpm16100531,
    author={Barmandah, Hassan and Bawazir, Omar A. and Marghalani, Siraj Aldeen and Shaat, Muath and Alkattan, Abdullah N. and AlEissa, Mariam M.},
    title={Beyond Mean Absolute Error: Ancestry-Stratified Calibration and Explainability for Warfarin Dosing Models},
    journal={Journal of Personalized Medicine},
    volume={16},
    year={2026},
    number={10},
    article-number={531},
    url={https://www.mdpi.com/2075-4426/16/10/531},
    issn={2075-4426},
    doi={10.3390/jpm16100531}
}

---

📄 License

This project is licensed under the Apache 2.0 License. This model was fit on IWPC data; the fitted artifact and code are shared under Apache-2.0, but the underlying IWPC dataset is not redistributed by this repository; see HasanBGit/WarfaRisk's data/DATA.md for how to obtain it.