P1853Comparative evaluation of risk model performance for prediction of 30-day mortality in transcatheter aortic valve replacement

Abstract

Abstract Background and purpose EuroSCORE (ES) and Society of Thoracic Surgeons (STS) risk prediction models are routinely used to guide decision-making for transcatheter aortic valve replacement (TAVR), however their accuracy remains limited, especially in very old and high-risk patients. New and updated scoring models have thus been developed to improve risk stratification. We performed a comparative evaluation of classical and new risk scoring models for prediction of 30d mortality in transcatheter aortic valve interventions. Methods and results A total of 1,569 patients undergoing transfemoral (TF, n=1.235) or transapical (TA, n=334) TAVR from 2009 to 2018 were included in a single-center all-comer analysis. Six risk scoring models (logES_I, ES_II, STS-PROM, FRANCE-2, OBSERVANT, GAV-2) were calculated for all patients and evaluated for prediction of 30d mortality in their model discrimination (c-indices with 95% confidence intervals (CI)) and calibration (graphical evaluation). Mean classical risk scores confirmed an intermediate-to-high-risk patient collective (logES_I 27.0±16.9%; STS-PROM 7.0±6.4%), mean 30-day mortality was 3.4% (TF 2.3%; TA 7.8%). Overall discrimination performance was best in FRANCE-2 (c-index 0.73, 95% CI 0.67–0.80), followed by STS-PROM (c-index 0.68, 95% CI 0.62–0.75), OBSERVANT (c-index 0.68, 95% CI 0.61–0.76), ES_II (c-index 0.64) and logES_I and GAV-2 (both c-indices 0.63). FRANCE-2 discriminated best in TF TAVR (c-index 0.72; range of c-indices 0.63 to 0.72), while OBSERVANT performed best in TA TAVR (c-index 0.70; range of c-indices 0.61 to 0.70). All risk scoring models – with the exception of lowest-risk deciles of STS-PROM and ES_II – showed an overestimation of mortality probability in all risk strata. Conclusion FRANCE-2 and OBSERVANT risk models showed superior discrimination performance to classical risk scoring models in TF and TA TAVR, however all models tended to overestimate mortality probability.