Methods/Statistics
A State-Wide Assessment of Dynamic Prediction Modeling Strategies for 30-day Postoperative Mortality in Transcatheter Aortic Valve Replacement (TAVR) Jackie Pollack* Jackie Pollack Wei Yang George Arnaoutakis Michael Kallan Stephen Kimmel
Background: With the increasing use of TAVR for severe aortic stenosis, robust mortality prediction models in this population are crucial. Early reliance on surgical aortic valve replacement (SAVR) models may inadequately capture evolving TAVR patient risks and differences in patient selection for the procedure. Deriving optimal TAVR-specific models initially faced challenges due to small sample sizes and the rapidly evolving nature of TAVR patients. One innovative strategy involves dynamically updating existing SAVR models for TAVR patients.
Methods: A baseline model, using all 14,070 surgical procedures (557 deaths) in Pennsylvania from 1999-2006, was developed to predict 30-day postoperative mortality in SAVR patients. In the subsequent test data (2012-2018, n=21,083 SAVRs, 514 deaths, and n=13,247 TAVRs, 348 deaths), we evaluated three distinct modeling strategies: 1) a static, non-updating approach, 2) updating the model in a combined SAVR+TAVR population, and 3) only updating the model in TAVR patients. Calibration regression (CR), which annually recalibrates the model intercept and slope, was applied to scenarios 2 and 3. Model performance was assessed in TAVR-only patients through measures of calibration and discrimination.
Results: TAVR mortality significantly decreased (from 4.9%-1.7%, p=0.016), while SAVR mortality remained stable (from 2.7%-2.0%, p=0.55). The static model consistently overpredicted the risk of TAVR mortality, with low discrimination (c-statistic (c-stat)=0.629). In contrast, both CR scenarios exhibited improvements in calibration (see figure) and modest discrimination improvement (combined population c-stat=0.651 and TAVR only 0.649, respectively) in TAVR patients.
Discussion: While waiting for ample TAVR data to facilitate the development of robust prediction models, the interim use of dynamic model updating is preferable over static methods. Yet, there is a need to develop and update new TAVR-specific models.
