Investigation of a Support Vector Model to Predict Lung Radiotherapy (RT) Induced Pneumonitis (RP)

Article Outline

• Purpose/Objective(s)
• Materials/Methods
• Results
• Conclusions
• Acknowledgement
• Copyright

Back to Article Outline

Purpose/Objective(s) 

To build and test a Support Vector Machine (SVM) model to predict the occurrence of Grade 2+ RT induced pneumonitis. SVM is a sophisticated statistical technique that uses a flexible hypersurface boundary to separate the cases with and without pneumonitis. Despite the flexibility, SVM is only minimally influenced by outliers that can destroy the predictive accuracy of other commonly used methods.

Back to Article Outline

Materials/Methods 

Two SVM models were built using data from 235 patients with lung cancer treated using RT (34 diagnosed with RP). One model (SVM_all) selected input features from all lung dose and non-dose factors. For comparison, the other model (SVM_dose) selected input features only from lung dose factors. The models were built with in-house developed software that employed a unique strategy to sequentially add/substitute features. During SVM model training, the data were randomly split into ten groups of approximately equal size. Nine groups were used to train the SVM and the remaining group was used as a test to measure model generalization capability. Using this methodology, each of the 10 groups was considered, in turn, as the test group (10-fold cross-validation). After the model was built, the significance of each input feature was individually evaluated by omitting it during SVM training and gauging its impact by the consequent deterioration in the cross-validated Receiver Operating Characteristics (ROC) area.

Back to Article Outline

Results 

The input features selected to build SVM_all were lung generalized equivalent uniform doses (EUD) with exponents a ≈ 1, chemotherapy prior to RT, central/peripheral tumor location, gender, and histology (adenocarcinoma/other; small cell/other). The input features for SVM_dose were EUD a ≈ 1, lung volume receiving >48 Gy (V₄₈), and V₅₀. Both models selected EUD a ≈ 1 (EUD a = 1 is the mean lung dose, which frequently appears as a strong predictor of RP in literature). The area under the cross-validated SVM_all ROC curve (Figure 1a) was 0.76 (sensitivity/specificity = 74%/75%). Compared to the corresponding SVM_dose area of 0.71 (sensitivity/specificity = 68%/68%) (Figure 1b), SVM_all was superior, indicating that non-dose features are important contributors to separating patients with and without RP. Among the input features selected by model SVM_all, only the dose factors (EUD with exponents a ≈ 1) were individually significant (p < 0.05).

Back to Article Outline

Conclusions 

The SVM model constructed from dose and non-dose factors is a powerful, yet robust, prospective tool for predicting the occurrence of radiation-induced lung pneumonitis.

Back to Article Outline

Acknowledgement 

Grants NIH R01 CA 115748 and NIH R01 CA69579.

• 
  • View Large Image
  • Download to PowerPoint