18P A radiomic approach to differentiate the immunotherapy-induced pneumonitis in patients with stage IV NSCLC

BackgroundImmunotherapy Induced Pneumonitis (IIP) is a rare lethal side effect of immune checkpoint inhibitors in patients with stage IV non-small-cell lung cancer. Timely diagnosis of IIP is crucial to treat the patients with high doses of corticosteroids and reduce the risk of death. To the best of our knowledge, only one study has been done on radiomics to detect IIP. However, the data used were limited, and their result cannot be conclusive. This study aimed to investigate whether the radiomic features are effective in detecting patients with IIP.MethodsIn a prospective clinical trial, 450 patients with stage IV NSCLC were recruited from six centres in the Netherlands and Belgium. The computed tomography images were obtained during the 6-week follow-up visits after immunotherapy. The validation of the model has been done using just CT scans acquired from the UMC Centre. Instead, the training data comprises the CT scans acquired on all the other centres. For the radiomics extraction, a lung mask for each CT scan has been segmented using the HU value of the images. The intraclass correlation coefficient was used to assess the reliability of the features. The features that showed poor reliability have been discarded (ICC<0.80).The recursive feature selection was used as the dimensionality reduction technique. To address the multicollinearity, the features with the Spearman correlation coefficient of higher than 0.75 were removed. The reliable radiomic features were fitted into a Logistic Regression classifier. The discrimination power was evaluated using the AUC ROC.ResultsA total of 806 out of 837 radiomics features were reliable. The recursive feature selection resulted in 403 features. After collinearity inspection, 42 features were selected. The AUC was 0.91 (95% CI: 0.75 to 0.98).ConclusionsRadiomic features have shown the potential to detect patients with IIP. Although the model showed good discriminative power, further investigations are needed to validate the proposed solution for clinical use.Legal entity responsible for the studyThe authors.FundingMaastro Clinic, Clinical Data Science Group.DisclosureAll authors have declared no conflicts of interest. BackgroundImmunotherapy Induced Pneumonitis (IIP) is a rare lethal side effect of immune checkpoint inhibitors in patients with stage IV non-small-cell lung cancer. Timely diagnosis of IIP is crucial to treat the patients with high doses of corticosteroids and reduce the risk of death. To the best of our knowledge, only one study has been done on radiomics to detect IIP. However, the data used were limited, and their result cannot be conclusive. This study aimed to investigate whether the radiomic features are effective in detecting patients with IIP. Immunotherapy Induced Pneumonitis (IIP) is a rare lethal side effect of immune checkpoint inhibitors in patients with stage IV non-small-cell lung cancer. Timely diagnosis of IIP is crucial to treat the patients with high doses of corticosteroids and reduce the risk of death. To the best of our knowledge, only one study has been done on radiomics to detect IIP. However, the data used were limited, and their result cannot be conclusive. This study aimed to investigate whether the radiomic features are effective in detecting patients with IIP. MethodsIn a prospective clinical trial, 450 patients with stage IV NSCLC were recruited from six centres in the Netherlands and Belgium. The computed tomography images were obtained during the 6-week follow-up visits after immunotherapy. The validation of the model has been done using just CT scans acquired from the UMC Centre. Instead, the training data comprises the CT scans acquired on all the other centres. For the radiomics extraction, a lung mask for each CT scan has been segmented using the HU value of the images. The intraclass correlation coefficient was used to assess the reliability of the features. The features that showed poor reliability have been discarded (ICC<0.80).The recursive feature selection was used as the dimensionality reduction technique. To address the multicollinearity, the features with the Spearman correlation coefficient of higher than 0.75 were removed. The reliable radiomic features were fitted into a Logistic Regression classifier. The discrimination power was evaluated using the AUC ROC. In a prospective clinical trial, 450 patients with stage IV NSCLC were recruited from six centres in the Netherlands and Belgium. The computed tomography images were obtained during the 6-week follow-up visits after immunotherapy. The validation of the model has been done using just CT scans acquired from the UMC Centre. Instead, the training data comprises the CT scans acquired on all the other centres. For the radiomics extraction, a lung mask for each CT scan has been segmented using the HU value of the images. The intraclass correlation coefficient was used to assess the reliability of the features. The features that showed poor reliability have been discarded (ICC<0.80). The recursive feature selection was used as the dimensionality reduction technique. To address the multicollinearity, the features with the Spearman correlation coefficient of higher than 0.75 were removed. The reliable radiomic features were fitted into a Logistic Regression classifier. The discrimination power was evaluated using the AUC ROC. ResultsA total of 806 out of 837 radiomics features were reliable. The recursive feature selection resulted in 403 features. After collinearity inspection, 42 features were selected. The AUC was 0.91 (95% CI: 0.75 to 0.98). A total of 806 out of 837 radiomics features were reliable. The recursive feature selection resulted in 403 features. After collinearity inspection, 42 features were selected. The AUC was 0.91 (95% CI: 0.75 to 0.98). ConclusionsRadiomic features have shown the potential to detect patients with IIP. Although the model showed good discriminative power, further investigations are needed to validate the proposed solution for clinical use. Radiomic features have shown the potential to detect patients with IIP. Although the model showed good discriminative power, further investigations are needed to validate the proposed solution for clinical use.