Abstract
PURPOSE: Cardiotoxicity is one of the major concerns in breast cancer treatment, significantly affecting patient outcomes. To improve the likelihood of favorable outcomes for breast cancer survivors, it is essential to carefully balance the potential advantages of treatment methods with the risks of harm to healthy tissues, including the heart. There is currently a lack of comprehensive, data-driven evidence on effective risk stratification strategies. The aim of this study is to investigate the prediction of cardiotoxicity using machine learning methods combined with radiomics, clinical, and dosimetric features. MATERIALS AND METHODS: A cohort of 83 left-sided breast cancer patients without a history of cardiac disease was examined. Two- and three-dimensional echocardiography were performed before and after 6 months of treatment to evaluate cardiotoxicity. Cardiac dose-volume histograms, demographic data, echocardiographic parameters, and ultrasound imaging radiomics features were collected for all patients. Toxicity modeling was developed with three feature selection methods and five classifiers in four separate groups (Dosimetric, Dosimetric + Demographic, Dosimetric + Demographic + Clinical, and Dosimetric + Demographic + Clinical + Imaging). The prediction performance of the models was validated using five-fold cross-validation and evaluated by AUCs. RESULTS: 58% of patients showed cardiotoxicity 6 months after treatment. Mean left ventricular ejection fraction and Global longitudinal strain decreased significantly compared to pre-treatment (p-value < 0.001). After feature selection and prediction modeling, the Dosimetric, Dosimetric + Demographic, Dosimetric + Demographic + Clinical, Dosimetric + Demographic + Clinical + Imaging models showed prediction performance (AUC) up to 73%, 75%, 85%, and 97%, respectively. CONCLUSION: Incorporating clinical and imaging features along with dose descriptors are beneficial for predicting cardiotoxicity after radiotherapy.