Abstract
Cardiotoxicity is the loss of the heart muscle's ability to contract effectively, often due to chemotherapy or radiation therapy. This study uses interpretable machine learning to predict post-chemotherapy cardiotoxicity using radiomics features extracted from the baseline echocardiography images. The study involved 100 cancer patients at Rajaei Cardiovascular Medical and Research Center. Baseline echocardiography images were used to extract radiomics features such as the left ventricular ejection fraction. According to the 12-month follow-up echocardiography, cardiotoxicity is defined as EF decline. Machine learning models predicted chemotherapy-induced cardiotoxicity based on radiomics features, with their significance confirmed via SHapley Additive exPlanations (SHAP) and Permutation-based Feature Importance Test (PermFIT). Among the 100 patients with a mean age of 54.5 ± 13.7, 41 patients (41%) experienced cardiotoxicity. For the short-axis view, the K-nearest neighbors (KNN) and Linear Support Vector Machine (SVM) models achieved accuracies of 0.92 and 0.90, respectively, with the best outcome of 92%. For the 4-chamber view, SVM and KNN reached accuracies of 0.88 and 0.83. These findings underscore the potential of machine learning, especially using short-axis echocardiography, to enhance early diagnosis of cardiotoxicity in chemotherapy patients.