Predicting heart dose in left-sided breast cancer patients using volumetric modulated arc therapy: an anatomical feature-driven machine learning model

Abstract The purpose of this study was to establish a machine-learning model that predicts heart dose in left-sided breast cancer patients treated with volumetric modulated arc therapy (VMAT). As radiotherapy (RT) poses an increased risk of cardiac toxicity, the model employs anatomical features to predict heart dose, tackling a significant issue in the management of breast cancer. This retrospective analysis focused on 53 patients with left-sided breast cancer who received VMAT RT. Various partial arc VMAT techniques were assessed, including the 2P, 4P and 5P methods. Key anatomical parameters measured included mean heart distance (MHD), total heart volume (THV) within the treatment field, heart volume (HV) and planning target volume (PTV). Elastic Net regression models were created to forecast heart dose metrics associated with different VMAT techniques. The Elastic Net regression models successfully predicted heart dose metrics, with VMAT-4P achieving the best performance, reflected in the lowest root mean squared error (RMSE) of 0·9099 and a median absolute error (MEDAE) of 0·5760 for the mean dose. VMAT-5P was particularly effective in predicting V5Gy, with an RMSE of 4·8242 and a MEDAE of 2·1188, while VMAT-2P recorded the lowest MEDAE for V25Gy at 1·0053. The feature importance analysis highlighted MHD as the primary predictor, contributing 75%, followed by THV at 18%, HV at 4% and PTV at 3%. The findings of this study emphasise the critical need to consider patient-specific anatomical features and the effectiveness of VMAT techniques in the treatment planning for left-sided breast cancer. The predictive models established present a pathway for personalised treatment enhancement. Treatment planners are encouraged to assess a range of anatomical characteristics when choosing the optimal VMAT technique.