β-Thalassemia · reports & figures
Model figures · bthal
Held-out synthetic test split · produced by
ml/bthal/evaluate.py.F1 · calibration
Reliability · isotonic
Research preview
Held-out synthetic test split.
F2 · decision curve
Net benefit vs threshold
Research preview
Benefit head · Vickers net benefit on the test split.
F3 · feature importance
SHAP — top features (signed)
Research preview
Mean |SHAP| across the test set, signed by feature–SHAP correlation. Positive bars push toward the benefit endpoint. Source: figures/shap_top20.json.
Illustrative · ancestry PCA
PC1 vs PC2
Research preview
Illustrative — not a model output; population structure is not used by the fitted model.
Illustrative · Fst
Pairwise Fst
Research preview
Thai-KMUTNB
Thai-Siriraj
Han-Chinese
Italian
AsiaPacific
EUR-ref
AFR-ref
Thai-KMUTNB
·
0.00
0.02
0.09
0.03
0.11
0.16
Thai-Siriraj
0.00
·
0.02
0.09
0.03
0.11
0.17
Han-Chinese
0.02
0.02
·
0.09
0.02
0.10
0.16
Italian
0.09
0.09
0.09
·
0.09
0.01
0.15
AsiaPacific
0.03
0.03
0.02
0.09
·
0.10
0.15
EUR-ref
0.11
0.11
0.10
0.01
0.10
·
0.15
AFR-ref
0.16
0.17
0.16
0.15
0.15
0.15
·
Fst00.08≥0.17
Illustrative — decorative genomics panel.
Comparator · conditioning regimens
Conditioning · Treosulfan-flu-thio
Comparator A
Research preview
2-yr OS
Second-HSCT incidence
- · 2-yr OS 94.7% · second-HSCT 9.0%
- · Lüftinger R et al., Ann Hematol 2022;101:655–665 (PMID 34999929)
- · n = 772 thalassemia patients · EBMT registry
Conditioning · Busulfan-flu
Comparator B
Research preview
2-yr OS
Second-HSCT incidence
- · 2-yr OS 92.7% · second-HSCT 4.6%
- · Lüftinger R et al., Ann Hematol 2022;101:655–665 (PMID 34999929)
- · n = 772 thalassemia patients · EBMT registry
Endpoint criteria
Locked definitions per probabilistic output
Research preview
loading endpoint definitions…