Exploration Of Establishing A Characteristic Model Of Immune Checkpoint Inhibitor-related Pneumonitis Based On Radiomics And Clinical Parameters

Objective:Based on the chest CT images of lung cancer patients,the radiomics features related to immune checkpoint inhibitor‐related pneumonitis were screened,and the characteristic model of immune checkpoint inhibitor‐related pneumonitis was constructed by machine learning and combined with clinical parameters,and distinguish with radiation pneumonitis and viral pneumonitis.Methods:Imaging data(chest CT images)and clinical parameters(basic clinical features and laboratory indicators)of lung cancer patients diagnosed with immune checkpoint inhibitor‐related pneumonitis(CIP)and radiation pneumonitis(RP),non-tumor patients diagnosed with viral pneumonitis(VP)in our hospital from October 2018 to December2022 were retrospectively collected.Firstly,statistical methods were used to analyze the clinical parameters of different types of pneumonitis,and the characteristics of clinical parameters related to various types of pneumonitis were screened.Then,the Py Radiomics toolkit was used to extract the radiomics features of the pneumonitis region in the CT images,and least absolute shrinkage and selection operator(LASSO)was used to further screen the radiomics features,and the 5 fold cross validation was performed,so as to obtain the image feature data highly relevant to the identification of various types of pneumonitis.Then four machine learning algorithms,random forest(RF),support vector machine(SVM),logistic regression(LR)and extreme gradient boosting(XGBoost),were used to establish the models of immune checkpoint inhibitor‐related pneumonitis,radiation pneumonitis and viral pneumonitis for the selected individual CT image features.In addition,the four machine learning algorithms were used again to establish models of immune checkpoint inhibitor‐related pneumonitis,radiation pneumonitis and viral pneumonitis for radiomics features of combined clinical parameters.Finally,receiver operating curve(ROC)and area under the curve(AUC)were used to evaluate the discrimination performance of these models.Results:1.A total of 166 patients meeting the inclusion criteria in this study,including 34cases CIP,62 cases RP and 70 cases VP.2.Basic clinical features of patients were analyzed,and theχ~2 showed that age,BMI,ECOG score and previous smoking in CIP vs RP,CIP vs VP and CIP vs VP of no statistical difference(P>0.05).There was no significant difference in sex between RP vs VP(P=0.353).There were statistically significant differences in CIP vs RP and CIP vs VP(P<0.05),and the male proportion in CIP group was higher than that in RP group(P=0.001),The male proportion of CIP group was higher than VP group(P=0.009).3.Kruskal-Wallis test results showed that there were 9 laboratory indicators related to the differentiation of three types of pneumonitis.Lactate dehydrogenase(LDH)in CIP group was higher than that in RP group,albumin(ALB)and platelet to lymphocyte ratio(PLR)were lower than that in RP group(P<0.05).CIP group had higher eosinophil count(EC),basophil count(BC)and ALB than VP group,but lower LDH than VP group(P<0.05).EC,BC and ALB of RP group were higher than those of VP group,and white blood cell count(WBC),absolute neutrophil count(ANC)and LDH of RP group were lower than those of VP group(P<0.05).Among which ALB and LDH were significant in the identification of pneumonitis in every two of the three groups.4.After Py Radiomics software package extraction and LASSO screening,a total of 26 different CT image features were obtained,namely 8 intensity features,7 shape features and 11 texture features.The results of modeling the individual radiomics features using four machine learning algorithms showed that the RF algorithm performed best,with an AUC of 0.662 for CIP vs RP,0.708 for CIP vs VP,and 0.777for RP vs VP(AUC>0.5 indicated that the model has the ability of discrimination).5.The results of modeling the radiomics features of the combined clinical parameters by using four machine learning algorithms showed that the RF algorithm performed best,with the AUC of 0.899 for CIP vs RP,0.85 for CIP vs VP,and 0.896for RP vs VP.Conclusions:1.Albumin and lactate dehydrogenase have clinical significance in differentiating immune checkpoint inhibitor‐related pneumonitis,radioactive pneumonitis and viral pneumonitis.2.The characteristic model of immune checkpoint inhibitor‐related pneumonitis can be established based on radiomics,and it can be distinguished from radiation pneumonitis and viral pneumonitis.3.The model of immune checkpoint inhibitor‐related pneumonitis established by combination of radiomics features and clinical parameters can further improve the differential efficacy of the model from radiation pneumonitis and viral pneumonitis.