Predicting acute radiation-induced toxicity in lung cancer patients undergoing radiotherapy on the basis of pre-treatment CT scans and dose distributions using machine learning

Abstract

Purpose: In this study, we used machine learning to predict acute radiation-induced toxicity for lung cancer patients undergoing radiotherapy on the basis of pre-treatment CT scans, contouring of organs, and radiotherapy dose distributions. The radiation oncologists can then use the acute toxicity prediction to create a treatment plan best suited for a patient.

Methodology: We prepared a data set that is larger than any other data set found in related work predicting radiation-induced toxicity in lung cancer patients. This data set consists of clinical features, dosimetrics, radiomics, and acute toxicity labels of 458 patients. The preprocessing of the data set is described in detail, which is essential for reproduction in future studies. Three classification algorithms were used in this study: Logistic Regression (LR), Random Forest (RF), and Support Vector Machine (SVM). Ridge regularization was used for both LR and SVM. The outcome was divided into four classes: any toxicity, fatigue, esophagus toxicity, and lung toxicity. For each outcome class, only dosimetrics and radiomics of organs related to the type of acute toxicity were used. A model was trained and evaluated for every combination of input feature type, classification algorithm, and output class. A nested 5-fold cross-validation was used to train the models and a grid search was performed to train the hyperparameters of each classification algorithm.

Results/conclusion: The best AUC scores for the outcome classes any toxicity, fatigue, esophagus toxicity, and lung toxicity were 0.72, 0.65, 0.86, and 0.57, respectively. Although difficult to directly compare AUC values due to the usage of different data sets, the AUC score of 0.86 is so far unmatched by related studies. For all classes except lung toxicity, the best AUC score was achieved using a LR model. The lung toxicity class was best predicted using RF, but this was likely due a deficiency in the input features requiring a nonlinear solution, rather than due to the nature of the class itself. Dosimetrics were found to be the most predictive features among the input types used in this study, followed by clinical features which performed slightly worse overall. The shape and first order radiomics were found to have almost no effect on the prediction of acute toxicity. Overall, the methods presented in this study are promising for the future of predicting acute toxicity.