Combining Radiomics Features with Immune Response Biomarkers to Build a XGBoost Model to Predict Radiation Pneumonitis (RP) in Patients with Primary Lung Cancer

Abstract

<h3>Purpose/Objective(s)</h3> Patients are heterogenous in their responses to radiation lung damage. It has been reported mostly from western people that the computed tomography (CT) radiomics features have a potential to identify patients at high risk for radiation pneumonitis (RP) in Western people. The purpose of this study was to 1) validate the significance of radiomics features on RP prediction, 2) exam the differences in systemic level of immune checkpoint indoleamine 2,3-dioxygenase (IDO) in patients with RP and without RP, and 3) explore the performance of the extreme gradient boosting (XGBoost) of combining above factors, on grade 2 and above RP in Chinese patients with primary lung cancer. <h3>Materials/Methods</h3> Planning CT scans and blood of baseline and end of treatment from 43 patients treated for primary lung cancer with radiotherapy were collected. Grade 2 and above RP was defined as cough or short of breath need medication treatment during or at the end of radiotherapy. Radiomics features were extracted from lung-GTV volume in planning CT using python package open-source software. Serum kynurenine, tryptophan and kynurenine: tryptophan ratio, which is IDO systemic activity related biomarkers (IDO biomarker) were measured at pre-RT and end of RT. The relation between features [radiomics features and IDO biomarker] and RP. Finally, the radiomics features with p value smaller than 0.1 were used for modeling. Patients were randomly split into 80% for training and 20% for validation. Model was built with XGBoost in train dataset and was tested in independent test dataset. The model predictive ability was assessed using area under the receiver operating characteristic curve (AUC). <h3>Results</h3> Seven out of 43 (16.3%) patients presented grade 2 and above RP. A total of 109 radiomics features were extracted. A total of 31 features, including 6 first-order, 4 gray level co-occurrence matrix (GLCM), 5 gray level dependence matrix (GLDM), 5 gray level run length matrix (GLRLM), 3 gray level size zone matrix (GLSZM) and 8 shape features, were significantly different (p value ≤ 0.05) between patients with RP and without RP. The IDO biomarkers at pre-RT and end of RT seemed to be non-significant (p value: 0.46-0.81). 43 radiomics features with p values smaller than 0.1 were used for model building. AUC of in the training dataset was 0.86 [95% CI 0.75-1] and of test dataset was 0.75 [95% 0.5-1]. A model of combined IDO biomarkers and radiomics features to build model, The predictive AUC of the training dataset was 0.9 [0.75-1] and of test dataset was 0.75 [0.5-1]. AUC slightly improved on training set. <h3>Conclusion</h3> This study at some degree validated the significance of radiomics features extracted from planning CT on predicting grade 2 and above RP in primary lung cancer in Chinese patients which has not been reported previously. IDO biomarkers did seem help, but the model built with XGBoost approach improved the predictive ability. Study with larger number of patients and ideally from multicenters are needed to validate this finding.