Comparative validation of seven ESC/ERS-consistent risk stratification scores in a nationwide Pulmonary Arterial Hypertension Registry

Background The 2022 ESC/ERS guidelines recommend multiparametric risk assessment for pulmonary arterial hypertension (PAH). Yet there is no strictly defined operational method for calculating a risk score when registry data is incomplete. Different guideline-consistent approaches can therefore be constructed, but it remains unclear whether these perform similarly. Purpose To compare the prognostic performance and real-world applicability of seven prespecified ESC/ERS-consistent strategies for deriving a three-category risk score. Methods We analyzed baseline data of PAH treatment-naive patients from the Hellenic Pulmonary Hypertension Registry (2015–2025), a longitudinal multicenter PH registry. Seven candidate scores (S1–S7) were predefined by combining different eligibility rules (parameter-count versus domain-based requirements) with different aggregation approaches (parameter-level versus domain-level grading). The endpoint was all-cause mortality. Model performance was evaluated using time-dependent AUC, integrated AUC, Uno’s C and Harrell’s C for discrimination, Brier score for prediction error, and calibration slope plus observed–predicted comparisons for calibration. Internal validity was assessed using bootstrap cross-validation. Results The analysis population comprised 310 patients (median age 60y, 67.7% females). During a median follow-up of 5.2 years 101 deaths occurred. Eligibility for scoring varied considerably, from 111 patients fulfilling stringent domain requirements to the full cohort when minimal criteria were applied. Mean AUC across 1–5 years ranged from 0.52 to 0.66. Domain-based strategies consistently demonstrated higher discrimination than parameter-count approaches. Among all methods, the ≥2-domain score (S3) showed the most balanced performance profile, combining high discrimination (Uno’s C=0.616, 95%CI: 0.564–0.674; iAUC=0.662, 95%CI: 0.581–0.737), robust calibration (mean slope=0.058), low prediction error (mean Brier=0.161) and broad applicability (n=253). The tiered score (S7) achieved similarly strong discrimination but showed less homogeneous risk group distribution, while the guideline-strict score (S4) produced the lowest prediction error at the expense of applicability (n=111). Minimal-requirement strategies showed the weakest performance. Bootstrap validation demonstrated minimal optimism and calibration slopes near unity across models. Conclusions The prognostic performance of ESC/ERS risk stratification is influenced not only by which variables are included but also by the implementation of logical and domain-based criteria when data are incomplete. Domain-based implementations outperform simple parameter-count approaches, and a ≥2-domain strategy provides the most favorable balance between discrimination, calibration, and applicability in clinical practice.