Synthetic-data demo · PDPA consent & IRB approval are required before any real-data use.details
More soon

thai-bthal-predict · β-Thalassemia

v1.0.0-rc3 · model-lockedResearch preview
A
β-Thalassemia · model card

β-Thalassemia — Model Card

Version v1.0.0-rc3 · last updated 2026-08-15T09:14:00Z
Research preview
Intended use

Research-preview decision-support for an MDT discussing curative therapy (HSCT or gene therapy) in transfusion-dependent β-thalassemia. NOT for clinical use, NOT an eligibility verdict.

Input features
  • HBB genotype, α-thal status, HbF QTL dosages (BCL11A, HBS1L-MYB, Xmn1-HBG2)
  • GSTA1 busulfan PGx star alleles
  • HLA match grade · HLA AA-mismatch score
  • CHIP (donor + recipient)
  • Age, baseline HbF%, transfusion burden, years transfused
  • Liver iron content, cardiac T2*, pre-tx ferritin
Method
Fitted XGBoost (per-endpoint, calibrated) · heuristic fallback
Active provider: bthal-heuristic · kind heuristic-demo. Numbers come from an in-app sigmoid; no fitted model is loaded.
The real fitted model is trained and served OUTSIDE this app. Lovable does not fit the model.
Training data
Fitted gradient-boosted trees (XGBoost): one probability-calibrated classifier per probabilistic endpoint plus regressors for the continuous outputs, trained on a SYNTHETIC, literature-anchored cohort (n=4000; base rates pinned to published anchors — EBMT 2024, Strocchio 2024, Frick et al. JCO 2022, CLIMB/Locatelli). NOT real patient data. Served via the model API (VITE_BTHAL_MODEL_URL); if the service is unavailable the app transparently falls back to the in-app heuristic — the active model version is shown on the Predict page.
Performance
Internal held-out metrics on a synthetic test split (see /validation): AUROC ≈ TI@24mo 0.65, VOD 0.69, graft-failure 0.57, cGvHD 0.56–0.62; continuous R² 0.27–0.69 (HbF@12mo 0.69). Modest and honestly reported — NOT externally or prospectively validated (externalValidation: false).
Known limitations
  • Discrimination is modest and, for some endpoints (graft-failure, cGvHD-d180), near chance on the held-out synthetic test split.
  • Trial anchors are context for orientation, not the model's prediction for this patient.
  • No prospective validation; no calibration against a real Thai β-thalassemia cohort.
  • Confidence intervals are illustrative, not data-derived.
Criteria versions
  • TI: EBMT 2024
  • VOD: EBMT 2023 refined
  • Graft failure: CIBMTR
  • cGvHD: NIH 2014
This interface presents outputs of a research-preview prediction model. It is not a regulated decision-support tool. Predictions come from one calibrated XGBoost classifier per endpoint, fitted on a synthetic-labeled cohort (no PHI; labels anchored to published base rates) and evaluated on a held-out synthetic test split. Not pre-registered; not externally validated.
Research preview · not for clinical use · synthetic data · model v1.0.0-rc3
criteria versions