P5463Prediction of the risk of valve surgery and adverse events in patients with aortic stenosis: myocardial tissue characterization with radiomics

Abstract

Introduction: Timing of surgery is an important factor to avoid irreversible my-ocardial damage in patients with aortic stenosis (AS). As a result, there is a need for new imaging biomarkers to improve risk stratification and predict patient out-come. Recently, it has been postulated that AS can be considered also a disease of the stressed myocardium, with fibrosis on cardiac MRI (CMR) being an independent predictor of mortality. Radiomics is a novel method for extraction of quantitative features from medical images, relating image features to phenotyp-ing, diagnosis and treatment through predictive modelling. Purpose(s): Characterization of fibrosis patterns on CMR using a radiomics analy-sis and machine learning to evaluate the prognostic added value in patients with AS and to improve risk stratification. Method(s): This study included 146 patients (average age: 68.5, 29% female) who underwent CMR at baseline with different degrees of AS; 55% of patients un-derwent AVR or died during the follow-up period. Segmented myocardium and different segmented regions were used for calculation of radiomic features and case-specific features, related to myocardial thickness. A temporal validation was first performed on this dataset, where 25% of the samples were used as training data. Random permutation (n=1000) demonstrated the representability of the fi-nal test set. To evaluate the performance in external validation, a second dataset was used (n=100, average age: 70.3, 50% female). Univariate analysis was done using the concordance index (CI). Multivariate analysis implemented a random forest (RF), linear support vector machine (SVM) and a generalized linear model (GLM). Feature selection was done with mRMR. A model with clinical features was included to determine performance on prediction of AVR, including aortic jet velocity, high-sensitivity troponin-I and ECG-strain. Result(s): Feature extraction resulted in 5639 features. Univariate analysis demon-strated 49 features in the first dataset that were prognostic (False Detection Rate <0.05, CI>0.6). Performance of the models in temporal validation showed a CI of 0.53, 0.55 and 0.58 for respectively SVM, RF and GLM. The GLM with clinical features had the best performance (CI: 0.86). The figure shows the result with respect to the random permutation. External validation showed a CI of respectively 0.66, 0.66 and 0.48. Thus, in external validation the performance improved (respectively 20% and 25%) for more sophisticated models such as RF and SVM At [Figure Presented] Conclusion(s): To the best of our knowledge, this is the first study to demonstrate the application of radiomics in CMR for improved prediction of patient outcome. More sophisticated models reveal a better performance in our external validation In addition, the results confirm the previously identified risk score based upon clinical features. Larger studies are needed to explicitly demonstrate improved prediction of AVR or adverse events in patients with AS. Funding Acknowledgements: Dutch Heart Association.