Machine Learning-Based Model for Predicting Radiation Pneumonitis in Locally Advanced Non- Small Cell Lung Cancer Treated with IMRT-A Two-Centre Study

Authors:
Preprint
Publisher website
Google Scholar
Abstract
Aim: To build and externally validate machine learning-based models for radiation pneumonitis (RP) prediction in patients with locally advanced non-small cell lung cancer (LA-NSCLC) treated with intensity-modulated radiation therapy (IMRT) in the era of precision radiotherapy. Patients and Methods: In this two-center retrospective study, a total of 218 patients (131 in the training cohort, and 87 in the external validation cohort) with LA-NSCLC. All patients underwent primary IMRT with strict lung dose constraints. Pretreatment CT radiomics features were extracted and then generated radiomics score (Rad-score). The study factors included Rad-score, dose-volume parameters and clinical features. Based on the independent risk factors, three machine learning models (random forest, logistic regression and decision tree) were developed and validated for predicting RP. The predictive performances of the models were evaluated using area under receiver operating characteristic curve (AUC) and decision curve analysis (DCA). Results: Within both cohorts, the overwhelming majority of patients were safely treated with radiotherapy within known lungs dose constraints. PE,ILD,N2-N3, ipsilateral lung Rad-score and contralateral lung Rad-score were independent risk factors for RP (P＜0.05). The AUC of random forest model, logistic regression model and decision tree model were 0.938, 0.859 and 0.632 in the training cohort, and 0.885, 0.911 and 0.721 in the external validation cohort, respectively. The calibration curve and DCA demonstrated goodness-of-ft and improved benefits in random forest model. Conclusion: PE, ILD, N2-N3 and CT radiomics features of lungs were independent predictors of RP in the LA-NSCLC patients treated with IMRT. The model using random forest algorithm exhibited the best predictive accuracy, outperforming logistic regression and decision tree.

This publication has 47 references indexed in Scilit: