Development and external validation of a novel echocardiography-based AI model for predicting aortic stenosis progression
K Faierstein, E I Itelman, T C Caller, Y S Shapira, Y A Altman, B Y Yaakobi, R K Kornowski, A S Segev, E M Maor, L P PerlAbstract
Background
Aortic stenosis (AS) is the most common valvular heart disease in older adults and requires timely identification of patients who need intervention. Nevertheless, current surveillance strategies rely on fixed follow-up intervals that do not adequately reflect the highly variable rate of disease progression among individuals. As a result, some patients undergo unnecessary imaging, while others who progress rapidly may experience delays in referral and treatment. We therefore aimed to develop and externally validate a highly practical machine learning (ML) model capable of predicting progression from mild or moderate AS to severe AS across several clinically relevant time horizons.
Methods
We conducted a retrospective cohort study using structured echocardiographic report data obtained from two independent tertiary medical centers in Israel. Three internally developed ML models were created to predict progression at 1, 3, and 5 years. The models incorporated a set of 27 echocardiographic and demographic variables derived from routine clinical reports, without the need for additional imaging processing. Model performance was assessed using the area under the receiver operating characteristic curve (AUC) and the F1-score in both internal testing and external validation cohorts.
Results
The models demonstrated strong discriminative performance across datasets from both institutions. In the internal test set, the AUC values were 0.93 for prediction of 1-year progression, 0.89 for 3-year progression, and 0.91 for 5-year progression. External validation confirmed robust generalizability, with AUC values of 0.91 for 1-year, 0.88 for 3-year, and 0.87 for 5-year progression.
Conclusion
A machine learning model based solely on structured clinical data can accurately forecast progression of AS and maintains high performance across different institutions. This approach has the potential to shift current practice from rigid guideline-based surveillance toward a personalized, risk-based follow-up strategy for patients with AS. The tool enables surveillance on predicted risk at 1, 3, and 5 years.ROC curves