Using electrocardiogram to stratify long-term mortality risk and prognosis in TAVR patients

Abstract Background Transcatheter aortic valve replacement (TAVR) is an important therapeutic approach for improving long-term outcomes in patients with severe aortic stenosis (AS). However, effective risk stratification tools are still needed for better long-term management of these patients. Recently, we developed an artificial intelligence electrocardiogram (AI-ECG) algorithm to assess the long-term mortality risk in TAVR patients. Objective This study aims to evaluate the predictive value of AI-ECG-based long-term mortality risk stratification for overall mortality risk in TAVR patients. Methods This retrospective study included patients with severe AS who underwent TAVR at a hospital from September 2012 to December 2021 and had an ECG performed and archived within 2 weeks before the procedure. The patients were randomly (1:1) divided into two groups (training and testing groups). An AI algorithm was trained using the ECG data from the training group to assess long-term mortality risk post-TAVR. The trained algorithm was then applied to predict mortality risk in the testing group, categorizing patients into high-risk and low-risk groups. Results A total of 460 TAVR patients were included (mean age 75.5 years, 41.8% female). Using a deep learning model, ECG images from 230 patients in the training group were used for feature learning and classification. During the training phase, a binary cross-entropy loss function (BCELoss) was used to evaluate the difference between predicted probabilities and true labels, with model parameters updated through backpropagation to improve accuracy. In the testing phase, the AI-ECG algorithm predicted outcomes for another 230 patients, converting predicted probabilities into binary results, with an accuracy of AUC = 0.85. Of the patients, 20 (8.7%) were classified as high-risk for death, while 210 (91.3%) were classified as low-risk. During a median follow-up of 46.6 months, 20 (95.2%) patients in the high-risk group died, while 16 (7.6%) in the low-risk group died (log-rank p 0.001). Multivariable-adjusted survival analysis revealed that high-risk classification was independently associated with an increased risk of death (adjusted HR aHR: 10.62, 95% CI: 3.56–31.69, p 0.001). This result remained stable across sensitivity and subgroup analyses. Conclusion In patients with severe AS undergoing TAVR, AI-ECG-based long-term mortality risk stratification is independently associated with overall mortality.ROC analysis in testing group K-M analysis across risk group