Machine Learning Classifier Using Pretreatment Factors to Predict Radiation Pneumonitis Following Stereotactic Body Radiation Therapy for Lung Cancer

Abstract

Listen

Translate article

<h3>Purpose/Objective(s)</h3> Symptomatic radiation pneumonitis (RP) is a common cause of treatment related toxicity in patients receiving radiation therapy for lung cancer. Machine learning techniques have been utilized to determine predictors of RP, but most of these models have used dosimetric features to generate results. The identification of pretreatment factors related to the development of symptomatic RP and implementation of these factors into a machine learning model would be clinically useful to guide treatment decisions, risk estimation, and risk reduction of development of symptomatic RP. <h3>Materials/Methods</h3> We retrospectively reviewed data from 278 lung cancer patients that underwent stereotactic body radiation therapy (SBRT). Pretreatment factors including patient demographics, pulmonary function data, and tumor characteristics were previously analyzed for the development of symptomatic RP, defined as CTCAE v4.0 ≥ Grade 2. A programming environment was used to generate classification machine learning models based on this dataset. Models were tested for accuracy in classification, area-under-the-curve (AUC), and special attention was paid to the false negative rate. The performance of several different algorithms using various combinations of predictors was evaluated including Decision Trees, Discriminant Analysis, Logistic Regression, Naive Bayes Classifiers, Support Vector Machines, Nearest Neighbor Classifiers and Ensemble Classifiers. Each model utilized 5-fold cross-validation to prevent overfitting. <h3>Results</h3> Of the 278 patients receiving SBRT, 42 patients (15.1%) developed symptomatic RP. A RUSBoosted Trees Ensemble Classifier was found to be the most accurate model tested. Features identified as significant risk factors for symptomatic RP in prior univariate and multivariate analyses were found to be the most important predictors in the machine learning models tested. These features included prior radiation treatment to the thorax or lung resection, location of tumor in the right lower lobe, a higher T Stage and age. When this model was used to classify cases of symptomatic RP, it had an AUC of 0.86 and an accuracy of 87%, correctly predicting 36 out of 42 pneumonitis cases. The model had a false negative rate of 14% and false positive rate of 13%. <h3>Conclusion</h3> A machine learning model was developed to identify pretreatment factors predictive of symptomatic RP. The advantage of this model when compared to others with a similar goal is that it depends solely on pretreatment factors, and therefore could be applied in a clinical setting prior to radiation therapy. Next steps include testing with an independent dataset to confirm the model's accuracy and prospective studies to verify the model's clinical utility.