Transcatheter aortic valve replacement (TAVR) is the treatment option of choice in inoperable and high-risk patients with severe aortic valve stenosis [1, 2] and recently showed favorable outcomes also in intermediate-risk [3, 4] and low-risk patients [5, 6]. Patient and procedural characteristics determine risk for adverse clinical outcomes. Cardiovascular societies recommend clinical decision-making for TAVR within the Heart Team [1, 2], assisted by the use of statistical risk scoring models for individual patient risk stratification. Classical surgical risk models such as logistic EuroSCORE I (LogES I) [7, 8], its update EuroSCORE II (ES II) [9], or the Society of Thoracic Surgeons Predicted Risk of Mortality model (STS PROM) [10] are routinely used to gauge operative risk; however, they are known to overestimate mortality risk in TAVR [11‐13]. Several TAVR-specific risk assessment models have been developed from national registry databases to optimize the predictive performance in the context of TAVR [14‐16]. However, the optimal approach to risk prediction in TAVR still remains unknown, as their performance has been poor in validation studies [12, 17, 18].

We here aimed to evaluate the 30-day mortality prediction performance of six surgery- and TAVR-specific risk models (Table 1) in our German Rhine Transregio Aortic Diseases cohort of patients undergoing TAVR.

Patient risk stratification for coronary procedures is well established—ranging from the elective setting to acute coronary syndromes [22‐24]—but the risk of TAVR procedures is thus far considerably less predictable: Patients suffer from a multitude of interdependent comorbid conditions, especially frailty as an essential risk factor is difficult to classify [25, 26]. The expansion of TAVR from highest risk/inoperable patients into intermediate- and lower risk groups introduces bias. On the procedural side, there are clear associations of center volume [27], operator experience [28], and choice of approach (TA vs. TF) [29] with clinical outcomes.

Classical surgical risk models (LogES I, STS PROM, ES II) have well-known limitations in TAVR [13]. LogES I is not recommended for use in TAVR anymore (1), we included the model for historical comparisons. National TAVR-registries have provided platforms for the development of TAVR-specific risk models [14‐16], which were mostly developed in the “early years” of TAVR (Table 1). Resulting limitations in their performance have already been seen when applied outside of their original populations [30, 31], external validation studies with a similar design to our study were performed in the United Kingdom [12] and in the Netherlands [18]—and also retrieved disappointing results.

This analysis from the German Transregio Aortic Diseases cohort elucidates risk prediction in the whole era of TAVR, which has developed from an experimental procedure in 2008 to a routine and high-volume alternative to surgical aortic valve replacement in 2018. With steadily increasing operator experience and new-generation devices, procedure counts have dramatically increased and complication rates declined over the study period. The TF approach has become the routine access route.

Surgical risk models performed similar to expectations [13]: Discrimination analysis essentially showed similar performance of STS PROM [10], LogES I [7, 8] and ES II [9] models. None could even remotely match discrimination performance in their original surgical patient cohorts. Calibration analysis underlined the general overestimation of 30-day mortality risk—most pronounced in the LogES I model. Surgical models are thus confirmed to have severe limitations to judge TAVR risk.

However, dedicated TAVR-specific models also disappointed: FRANCE-2 (overall c-index 0.66, 95% CI 0.60–0.71) was better than in the original validation cohort (c-index 0.59 [15]) and in the UK (c-index 0.62 [12]) and Netherlands (c-index 0.63 [18]) external validation studies, but not superior to surgical models and slightly worse than in an Israeli external validation study (c-index 0.71 [30]). The model considerably overestimated risk (Fig. 2) in our patients. The GAVS-II model [16] derived from the German Aortic Valve Registry was expected to be most adapted to German TAVR conditions, but it performed similar to the other models in the overall cohort (c-index 0.63) and in TF and TA subgroups and could not meet the performance in the development cohort (c-index 0.74), similar to the Dutch external validation study [18]. Numerically, its discrimination improved over time and in new vs. old generation devices (Table 4). The model also overestimated mortality (expected: 9.4%; Fig. 2). Reasons for the lower-than-expected performance may lie in the GAVS-II model being developed in 55% surgical/45% TAVR patients from 2011–2012 (Table 1), which significantly differed in age (74 vs. 81 years) and comorbidities from our population [16]. The OBSERVANT model [14] discriminated numerically worst in the overall cohort and in TF/TA subgroups, and could not match performance in the original population (c-index 0.71), which confirms findings in other external validation studies [12, 18, 30]. However—of all analyzed models—its mean mortality prediction (5.8%) came closest to observed events (3.7%).

Taken together, this analysis clarifies that risk prediction in TAVR is still an unsolved issue: all tested models are severely limited in their performance, and dedicated TAVR-specific models are not superior to decade-old surgical scores. Using any of these models for risk stratification is better than a coin-flip—but with ample room for improvement. A recently published analysis from the Netherlands [18] found similar results to the UK study [12] and to our results, which supports the validity of these findings across Europe. With TAVR evermore becoming a routine procedure and 30-day mortality rates reaching all-time-lows in high-volume Heart Centers, a re-calibration of existing risk models or a new development from growing registries is thus mandatory to improve accuracy [32, 33]. High-quality TAVR databases should enable us to produce risk models with performance comparable to surgical procedures or acute coronary syndrome. Incorporation of functional status/frailty assessment as important predictive factors may additionally improve model accuracy in the TAVR setting [34, 35].

Three surgical as well as three TAVR-specific risk scoring models showed mediocre performance in prediction of 30-day mortality risk for TAVR in the German Rhine Transregio Aortic Diseases cohort. Development of new or updated risk models is necessary to improve risk stratification.