Robustness of Machine Learning Predictions for Determining Whether Deep Inspiration Breath-Hold Is Required in Breast...

Background/Objectives: Deep inspiration breath-hold (DIBH) is a commonly used technique to reduce the mean heart dose (MHD), which is critical for minimizing late cardiac side effects in breast cancer patients undergoing radiation therapy (RT). Although previous studies have explored the potential of machine learning (ML) to predict which patients might benefit from DIBH, none have rigorously assessed ML model performance across various MHD thresholds and parameter settings. This study aims to evaluate the robustness of ML models in predicting the need for DIBH across different clinical scenarios. Methods: Using data from 207 breast cancer patients treated with RT, we developed and tested ML models at three MHD cut-off values (240, 270, and 300 cGy), considering variations in the number of independent variables (three vs. six) and folds in the cross-validation (three, four, and five). Robustness was defined as achieving high F2 scores and low instability in predictive performance. Results: Our findings indicate that the decision tree (DT) model demonstrated consistently high robustness at 240 and 270 cGy, while the random forest model performed optimally at 300 cGy. At 240 cGy, a threshold critical to minimize late cardiac risks, the DT model exhibited stable predictive power, reducing the risk of overestimating DIBH necessity. Conclusions: These results suggest that the DT model, particularly at lower MHD thresholds, may be the most reliable for clinical applications. By providing a tool for targeted DIBH implementation, this model has the potential to enhance patient-specific treatment planning and improve clinical outcomes in RT.