Abstract
BACKGROUND: Transcatheter aortic valve implantation (TAVI) aims to improve symptoms and prognosis, while minimising adverse outcomes. Available prediction models focus on individual outcomes, but those combining adverse events and symptom improvement in a single prediction model are scarce, and include only few variables and lack external validation. Using machine learning, we developed a clinically relevant model to identify patients at high risk of both adverse events and poor symptom improvement after TAVI. METHODS: In total, 72 candidate variables including clinical, medication use, biomarkers and (AI-derived) echocardiographic parameters were collected in patients with severe symptomatic AS undergoing TAVI. The primary outcome was a combination of poor symptom improvement (NYHA compared with baseline) and a composite of cardiovascular mortality, stroke or heart failure hospitalisation) at one year follow-up. LASSO Logistic regression was used for variable selection. External validation was performed in the Bern TAVI-registry. RESULTS: From a total of 827 patients (age 79.2 (± 7.29), 53% female), 101 patients (12%) had both adverse events and poor symptom improvement during one-year follow-up after TAVI, while 529 (64%) improved without any adverse events. Predictors for the combined primary outcome were history of COPD, use of vitamin-K antagonist, concomitant heart failure, reflected by mineralocorticoid receptor antagonists use, lower sodium and higher urea and (log-)NT-proBNP levels, lower AV mean gradient and larger LVOT diameter (area under the curve (AUC): 0.74 (internal validation: 0.72 and external validation: 0.66)). CONCLUSION: Our externally validated model can reasonably identify patients with both poor symptom improvement and adverse events after TAVI.