Abstract
Background
Radiotherapy is a standard treatment option for breast cancer, but it may lead to significant late morbidity, including radiation heart damage. Breast irradiation performed individually in the supine or prone position may aid in minimizing the irradiation dose to the heart and LAD coronary artery. A series of CT scans and therapy plans are needed in both positions for the ‘gold standard’ decision on the preferable treatment position. This method is expensive with respect to technology and physician workload.
Our ultimate goal is to develop a predictive tool to identify the preferable treatment position using easily measurable patient characteristics. In this article, we describe the details of how model building and consequently validation of the best model are done.
Methods
Different models were used: both logistic regression and multiple linear regressions were used to estimate the LAD mean dose difference (the difference between the mean dose to the LAD in the supine position versus prone position); predicted dose differences were analysed compared to the ‘gold standard’ values, and the best model was selected accordingly. The final model was checked by random cross-validation. In addition to generally used measures (ROC and Brier score), decision curves were employed to evaluate the performance of the models.
Results
ROC analysis demonstrated that none of the predictors alone was satisfactory. Multiple logistic regression models and the linear regression model lead to high values of net benefit for a wide range of threshold probabilities. Multiple linear regression seemed to be the most useful model. We also present the results of the random cross-validation for this model (i.e. sensitivity of 80.7% and specificity of 87.5%).
Conclusions
Decision curves proved to be useful to evaluate our models. Our results indicate that any of the models could be implemented in clinical practice, but the linear regression model is the most useful model to facilitate the radiation treatment decision. In addition, it is in use in everyday practice in the Department of Oncotherapy, University of Szeged, Hungary.
An Additive Sparse Penalty for Variable Selection in High-Dimensional Linear Regression Model
