- 作者：Rainer J. Klement ; Michael Allg盲uer ; Steffen Appold ; Karin Dieckmann ; Iris Ernst ; Ute Ganswindt ; Richard Holy ; Ursula Nestle ; Meinhard Nevinny-Stickel ; Sabine Semrau ; Florian Sterzing ; Andrea Wittig ; Nicolaus Andratschke ; Matthias Guckenberger
- 刊名：International Journal of Radiation Oncology*Biology*Physics
- 出版年：1 March, 2014
- 年：2014
- 卷：88
- 期：3
- 页码：732-738
- 全文大小：362 K
Background
Several prognostic factors for local tumor control probability (TCP) after stereotactic body radiation therapy (SBRT) for early stage non-small cell lung cancer (NSCLC) have been described, but no attempts have been undertaken to explore whether a nonlinear combination of potential factors might synergistically improve the prediction of local control.Methods and Materials
We investigated a support vector machine (SVM) for predicting TCP in a cohort of 399 patients treated at 13 German and Austrian institutions. Among 7 potential input features for the SVM we selected those most important on the basis of forward feature selection, thereby evaluating classifier performance by using 10-fold cross-validation and computing the area under the ROC curve (AUC). The final SVM classifier was built by repeating the feature selection 10 times with different splitting of the data for cross-validation and finally choosing only those features that were selected at least 5 out of 10 times. It was compared with a multivariate logistic model that was built by forward feature selection.
Results
Local failure occurred in 12% of patients. Biologically effective dose (BED) at the isocenter (BEDISO) was the strongest predictor of TCP in the logistic model and also the most frequently selected input feature for the SVM. A bivariate logistic function of BEDISO and the pulmonary function indicator forced expiratory volume in 1聽second (FEV1) yielded the best description of the data but resulted in a significantly smaller AUC than the final SVM classifier with the input features BEDISO, age, baseline Karnofsky index, and FEV1 (0.696聽卤聽0.040 vs 0.789 卤 0.001, P<.03). The final SVM resulted in sensitivity and specificity of 67.0%聽卤聽0.5% and 78.7%聽卤聽0.3%, respectively.
Conclusions
These results confirm that machine learning techniques like SVMs can be successfully applied to predict treatment outcome after SBRT. Improvements over traditional TCP modeling are expected through a nonlinear combination of multiple features, eventually helping in the task of personalized treatment planning.