Clinically interpretable machine-learning model for predicting heart mean dose using simple BEV-based metrics
1Department of Radiation Physics, Kidwai Memorial Institute of Oncology, Bengaluru, Karnataka 560029, India.
Abstract:
Radiation-induced cardiac toxicity remains a major concern in left-sided breast cancer radiotherapy, with mean heart dose (MHD) serving as a key predictor of long-term cardiac morbidity. This study aimed to develop a clinically interpretable machine-learning model to predict MHD using simple beam's-eye-view (BEV)-based heart-projection metrics. A retrospective cohort of 127 patients treated with postoperative left breast or chest wall radiotherapy including supraclavicular nodal irradiation was analyzed. Heart projections in the lateral and vertical directions were measured from medial and lateral tangential fields, defined as medial tangential horizontal (MTH), lateral tangential horizontal (LTH), medial tangential vertical (MTV), and lateral tangential vertical (LTV). Outliers were removed using interquartile range criteria. A multivariable linear regression model was developed using 5-fold cross-validation to predict MHD. In addition, a logistic regression classifier was trained to categorize patients suitable for three-dimensional conformal radiotherapy (MHD ≤ 4 Gy) versus inverse planning (MHD > 4 Gy). Model performance was evaluated using R², root mean squared error (RMSE), mean absolute error (MAE), Pearson correlation coefficient, accuracy, and area under the receiver operating characteristic curve (AUC). The linear regression model demonstrated strong predictive performance during cross-validation (R² = 0.69, RMSE = 0.61 Gy, MAE = 0.47 Gy), with a significant correlation between predicted and measured MHD (r = 0.83, p < 0.001). Independent validation further improved performance (R² = 0.76, RMSE = 0.69 Gy, MAE = 0.56 Gy; r = 0.90, p < 0.001). The logistic regression classifier achieved an overall accuracy of 88% with excellent discrimination (AUC = 0.95). Simple beam's-eye-view (BEV)-based heart-projection metrics can reliably predict MHD, enabling rapid preplanning assessment and supporting early selection of cardiac-sparing radiotherapy techniques.
