Program Information
Predicting Distant Failure in Lung SBRT Using Clinical Parameters
Z Zhou*, N Cannon , M Folkerts , P Iyengar , H Choy , R Timmerman , S Jiang , J Wang , UT Southwestern Medical Center, Dallas, TX
Presentations
TH-AB-304-5 (Thursday, July 16, 2015) 7:30 AM - 9:30 AM Room: 304
Purpose:To predict distant failure in lung Stereotactic Body Radiation Therapy (SBRT) using clinical parameters by machine learning algorithms with a new optimal parameter selection strategy.
Methods:Support Vector Machine (SVM) based method is used to construct the model to predict distant failure using 143 lung cancer patients underwent SBRT at our institute. Meanwhile, Artificial Neural Network (ANN) and Logistic Regression (LR) based methods are applied for comparisons. The clinical parameters for each patient include demographic parameters (4), tumor characteristics (6), treatment faction schemes (2) and pretreatment medicines (6). The total number of clinical parameters is 18. The model construction method consists of two steps: (1) selecting optimal parameter set by a new Clonal Selection Algorithm (CSA); and (2) constructing SVM, ANN and LR based models for predicting distant failure in lung SBRT. Selecting optimal parameter set is considered as a combinatorial optimization problem. The proposed CSA method is one of the evolutionary algorithms that can obtain global optimal solutions, overcoming the limitation of the ranking based selection method that lead to local optimization. In addition, to overcome the effect of imbalanced data in training process, a Synthetic Minority Over-sampling Technique (SMOTE) is applied. Sensitivity, specificity and accuracy are used to evaluate the performance of the constructed models.
Results:The accuracies for ANN, LR and SVM are 70.59%, 72.04%, 75.47%, respectively. The sensitivities for ANN, LR and SVM are 76%, 60.47% and 82%, while the specificities for the three models are 58.14%, 69.77% and 60.47%.
Conclusion:A new CSA based parameter selection method is proposed. The SVM with the proposed CSA optimal parameter set selection strategy can achieve better performance than ANN and LR for predicting distant failure in lung SBRT patients based on clinical parameters.
Contact Email: