CUORE and SCORE2 Validation in an Italian Primary Care Cohort

Key Takeaways

• Both models showed moderate 10-year discrimination, with CUORE showing slightly higher AUC ranges than SCORE2.
• Broader administrative endpoint ascertainment was associated with higher observed event rates for both models without materially changing discrimination.
• SCORE2 appeared more stable in calibration, while CUORE had slopes below 1, and the authors concluded that transparent endpoint reporting and local calibration assessment were important before applying absolute risk thresholds.

In ATS Milan, a head-to-head external validation tested CUORE and SCORE2 for 10-year prediction of first fatal or non-fatal myocardial infarction or stroke.

The study followed 38,101 adults from an Italian primary care-administrative cohort in ATS Milan for 10 years through linked hospital discharge and mortality registries. Baseline data were collected between 2007 and 2009. The outcome was a first fatal or non-fatal myocardial infarction or stroke during follow-up. Validation used a strict administrative definition based on primary diagnosis or procedure fields and a broad definition based on any diagnosis or procedure field. That broader-versus-strict outcome mapping was built into the external validation framework.

Discrimination was assessed with inverse probability of censoring weighting time-dependent AUC over 10 years. SCORE2 ranged from 0.713 to 0.720, while CUORE ranged from 0.730 to 0.741. CUORE showed a modest numerical edge in distinguishing future event risk in this cohort. Across the validation analyses, both scores maintained moderate discrimination and robust ordinal risk stratification over 10 years.

Calibration was examined with decile plots, calibration-in-the-large, and calibration slope, focusing on absolute risk estimation rather than relative ranking. SCORE2 was more stable in calibration overall, whereas CUORE showed calibration slopes below 1. Broader ascertainment was associated with higher observed event rates, rising from 4.4% to 5.6% for SCORE2 and from 5.0% to 6.8% for CUORE. That divergence reflected agreement between predicted and observed risk rather than relative ranking. Those broader definitions did not materially change AUC, indicating that endpoint mapping mainly changed absolute risk estimates.

Uncalibrated and recalibrated implementations were compared, and absolute risk estimates were sensitive to both administrative endpoint mapping and recalibration strategy. Ordinal risk stratification remained robust, but absolute risk estimation varied across these analytic choices. In the CUORE and SCORE2 validation, the authors concluded that transparent endpoint reporting and local calibration assessment were important before applying absolute risk thresholds. Overall, ordinal ranking was stable while absolute risk estimation was more variable across analytic choices.