Machine learning identified two distinct clinical profiles associated with increased risk of death after TAVR in a cohort of 220 patients, of whom 18% died over a median 3.2-year follow-up.
Cohort (n=220)
Does machine learning-based clustering of pre-procedural variables identify distinct clinical profiles associated with increased mortality risk in patients undergoing TAVR?
Machine learning-based clustering identifies distinct, non-uniform clinical profiles that predict mortality after TAVR, highlighting the potential for personalized risk stratification.
Abstract Background Transcatheter aortic valve implantation (TAVR) is an effective treatment for patients with severe aortic stenosis. Despite a meticulous selection for this intervention, even 25% of patients die within 1 year after the procedure. Therefore, there is a need to accurately determine patient characteristics that may predict poor outcome after TAVR. Aim In this study, we used an unsupervised and supervised machine learning (ML) approach to delineate distinct patient categories with varying demographic, clinical and echocardiographic features to define specific profiles with heightened mortality risk. Methods A cohort of 220 patients scheduled for TAVR underwent the collection of pre-procedural data, which comprised echocardiographic evaluation with strain and myocardial work analysis, 6-minute walk test (6MWT), laboratory tests including NT-proBNP, and information on frailty and comorbidities. Results During a median follow-up of 3.2 years, 39 patients (18%) died. Based on the clinical relevance of ML analysis outputs, a 4-cluster categorization using the Partition Around Medoids (PAM) method was selected and applied to the 57 pre-procedure variables without using outcome data. Basic characteristics of clusters are presented in Table 1. Two of the phenogroups (clusters 3 and 4) were associated with increased risk of death after TAVR (Table 2). These 2 distinct profiles were characterized by: (1) male predominance with severe LV and right ventricular abnormalities and pulmonary circulation involvement, frequent coincidence of low-flow low gradient conditions and severely elevated NT-proBNP; and (2) oldest age, female predominance, highest burden of comorbidities, moderately increased NT-proBNP, intermediate impairment in LV systolic function, and severe abnormalities of LV diastolic function. In both of these clusters, exercise capacity (6MWT distance) was more impaired than in the other two clusters (1 and 2). Conclusions Clinical profiles associated with increased risk of death after TAVR are not uniform, which highlights the complexity of prognostic trajectories in aortic stenosis. ML-based clustering offers a more personalized approach to predicting post-TAVR mortality, thus facilitating patient selection for this procedure.
Kosmala et al. (Thu,) conducted a cohort in Severe aortic stenosis (n=220). Machine learning-derived clinical profiles (clusters 3 and 4) vs. Other clinical profiles (clusters 1 and 2) was evaluated on Death. Machine learning identified two distinct clinical profiles associated with increased risk of death after TAVR in a cohort of 220 patients, of whom 18% died over a median 3.2-year follow-up.