ASNet: mortality and valve replacement prediction in patients with aortic stenosis using explainable artificial intelligence in a multi-ethnic asian cohort
E H Lim, S Jauhari, C H Sia, R Cherian, W Kong, P Chai, J Yip, T C Yeo, K K Poh, W YeungAbstract
Background
Better risk stratification in aortic stenosis (AS) for prognostication and decision for early valvular intervention purposes is in need. Complex relationships between various clinical parameters for risk stratification may not be easily delineated using conventional linear regression methods. A machine learning algorithm, Extreme Gradient Boosting (XGBoost), is able to capture nonlinear relationships among parameters. Our study aims to utilise XGBoost to develop and validate an explainable model (ASNet) for predicting clinical outcomes in patients with AS using readily available clinical and echocardiographic data.
Methods
We conducted a retrospective, multicentre cohort study on 2759 patients with at least moderate aortic stenosis across three tertiary hospitals between January 2014 and February 2023. Moderate and severe aortic stenosis were defined as 1.0-1.5cm2 and <1.0cm2 respectively by the continuity equation on transthoracic echocardiography. The primary outcome was defined as a composite of all-cause mortality and aortic valve intervention within 3 years from the date of index echocardiography. Patients from two hospitals were used for the ASNet model development and internal validation while external validation was performed on patients from the third hospital. ASNet was developed using various clinical and echocardiographic parameters, evaluated using discrimination, calibration, decision curve analysis, SHAP-based interpretability with subgroup analysis.
Results
977 patients (35.4%) met the primary outcome with 641 deaths, 287 surgical valve replacements and 49 catheter-based interventions. The ASNet model demonstrated discrimination in the internal test cohort with area under the curve (AUC) 0.92 (F1-score 0.81–0.84) and maintained robust performance in the external validation cohort (AUC 0.83–0.84; F1-score 0.63–0.67) (Figure 1). SHAP analysis demonstrated that comorbidities of ischemic heart disease, hypertension, older age, echocardiographic parameters of lower left ventricular outflow tract stroke volume, higher aortic peak velocity were the strongest contributors to predicted risk of the primary outcome (Figure 2). Subgroup analyses have shown that ASNet maintained its performance in both moderate (AUC 0.815) and severe (AUC 0.818), elderly patients of age >75 years (AUC 0.04), low left ventricular ejection fraction (<40%) (AUC 0.842) and gender (male AUC 0.830; female AUC 0.846). ASNet risk stratified groups has also demonstrated poorer survival in the high-risk group in Kaplan-Meier survival analysis (log rank p<0.001).
Conclusion
ASNet is an explainable and externally validated model which is based on readily available clinical and echocardiographic data. It has demonstrated ability to predict clinical outcomes in patients with at least moderate AS. This may allow earlier identification of high-risk patients with aortic stenosis, enabling timely interventions or closer surveillance.ROC curveSHAP analysis plot