| Score | AUC (95% CI) | ΔAUC vs ESS | DeLong z | p-value | Result |
|---|---|---|---|---|---|
| EndoSysScore (ESS) | 0.785 (0.74–0.83) | reference | — | — | — |
| EuroSCORE II | 0.744 | +0.041 | 2.23 | 0.026 | WIN |
| EndoSCORE | 0.749 | +0.035 | 2.17 | 0.030 | WIN |
| RISK-E | 0.722 | +0.063 | 2.86 | 0.004 | WIN |
| AEPEI (5-var) | 0.724 | +0.060 | 2.68 | 0.007 | WIN |
| APORTEI | 0.721 | +0.063 | 3.09 | 0.002 | WIN |
| STS-IE | 0.695 | +0.090 | 3.60 | 0.0003 | WIN |
| Phenotype | Definition | n | Observed | ESS predicted | Gap ESS | Best comparator gap |
|---|---|---|---|---|---|---|
| A · Rare high-lethality | Multivalve + S.aureus + CKD≥3/dialysis/PAPs>50 | 11 | 54.5% | 46.2% | −8.3 pp | −12.5 pp |
| B · Rare high-lethality | Age≥80 + NVE + (COPD or LVEF<50%) | 15 | 53.3% | 46.3% | −7.1 pp | −22.2 pp |
| C · Rare high-lethality | Abscess/periannular + isolated mitral + age≥65 | 10 | 50.0% | 50.0% | 0.0 pp | −4.1 pp |
| D · Rare high-lethality | PVE + S.aureus + mitral + aortic | 10 | 50.0% | 46.5% | −3.5 pp | 0.0 pp |
| E · Rare high-lethality | NVE + age≥80 + LVEF<50% | 11 | 54.5% | 52.8% | −1.7 pp | −21.9 pp |
| F · Rare high-lethality | Abscess + isolated mitral + age≥65 | 10 | 50.0% | 50.0% | 0.0 pp | −4.1 pp |
| Pooled (distinct · E⊆B, F=C counted once) | 44 | 50.0% | 46.0% | −4.0 pp | MAD ESS 3.4 pp vs RISK-E 15.7 pp | |
EndoSysScore (ESS) is a stacking ensemble developed on the GIROC Italian multicentre registry (2000–2023; development cohort n=4,285; 23 centres; 30-day mortality 10.6%). Rare high-lethality phenotypes were identified in the derivation cohort by systematic multivariable pattern analysis and used to guide conditional synthetic data augmentation (CTGAN + TIMA clinician-supervised plausibility filtering). The final model combines three base learners (XGBoost shallow, XGBoost moderate, logistic regression) through a LightGBM meta-learner, with isotonic and beta calibration applied on real development data only.
Model performance was evaluated in a held-out internal validation cohort of 930 patients (109 deaths, 11.7% mortality). Development (n=4,285) and validation (n=930) cohorts were drawn from the same 23-centre GIROC registry via a stratified random patient-level split (≈80/20 within each centre); all 23 centres contributed to both cohorts, with no patient overlap. No validation-cohort patient was used for model training, synthetic augmentation, calibration, or phenotype optimisation. Fully external validation (leave-one-centre-out and prospective inter-registry evaluation) is the planned next step.
Phenotype transparency note: during phenotype definition, candidate phenotypes were required to demonstrate non-zero representation in the validation cohort to confirm applicability to rare IE presentations. This feasibility check verified representation only; phenotype definitions were fixed before any outcome data from the validation cohort were examined. Full methodology, reproducible code, and de-identified data are openly deposited on Zenodo and GitHub.