Background As immunotherapy is more widely used for advanced non-small cell lung cancer (NSCLC), important challenges remain with adverse events including checkpoint inhibitor-associated pneumonitis (CIP). CIP usually requires discontinuation of immunotherapy even if the tumor responds, and effective models for predicting CIP are still limited. This study aims to investigate radiomic features using artificial intelligence (AI) algorithms to predict CIP in patients with NSCLC.
Methods Data from 132 patients with stage III-IV NSCLC treated with immunotherapy were analyzed. Tumor response was evaluated based on immune-related RECIST (irRECIST) (tables 1 and 2). Patients were categorized into two groups: durable responders (including complete response [CR], partial response [PR], or stable disease [SD]) and non-responders (progressive disease [PD]). Segmentation was performed by three physicians using LIFEx software (IMIV/CEA, Orsay, France). 3D-radiomic features were collected from the tumor and peritumoral regions on contrast enhanced CT imaging. The lesion size was measured as a volume-of-Interest (VOI). Linear mixed-effects (LME) regression model was used to evaluate the association between radiomic features and the size of the VOI, with CIP and radiation-related pneumonitis status specific variables for slope and intercept. The chemotherapy containing regimen was considered as a random factor. The corresponding p-values were used to assess differences in LME regression slope and/or intercept between all groups.
Results CIP was more common in the non-responder group [durable responder: 8/91 (8.8%) vs non-responder: 6/41 (14.6%) (p=0.049)] (table 2). Higher platelet counts at baseline (≥400K) was more prevalent in the non-responder group (p<0.001) and patients who received chemotherapy containing regimen were more likely to be durable responders (p<0.001) (table 1).
Association between some CT-based radiomic features and the tumor size demonstrates statistically significant difference between radiation-related and CIP. Significant difference in the intercept (0.05, 95%CI[0.04, 0.05] p<0.001) and the slope (-0.01, 95%CI[-0.007, 0.012], p<0.001) of Neighboring Gray Tone Difference Matrix (NGTDM) coarseness computed for primary tumors was found with CIP (intercept difference -0.07, 95%CI[-0.1, -0.4], p<0.001 and slope difference 0.02, 95%CI[0.01, 0.02], p<0.001). For the peritumoral space, Intensity Histogram Maximum Histogram Gradient Grey Level, also showed significant difference in the intercept (30, 95%CI [19.7–40.3], p<0.001) and the slope (-2.96, 95%CI[-5.3, -0.6], p<0.015) and CIP associated difference in the intercept equal 42.7 (95%CI[12.76, -70.4], p<0.03) and slope difference -10.5 (95%CI[-15.3, -4.76], p<0.002).
Conclusions Non-responders among NSCLC patients treated with immunotherapy had a higher incidence of CIP, and CT-based radiomic features may assist in the prediction of CIP.
Ethics Approval Northwestern IRB approved: STU00207117EnterWrite to Liam Il-Young Chung
This is an open access article distributed in accordance with the Creative Commons Attribution Non Commercial (CC BY-NC 4.0) license, which permits others to distribute, remix, adapt, build upon this work non-commercially, and license their derivative works on different terms, provided the original work is properly cited, appropriate credit is given, any changes made indicated, and the use is non-commercial. See http://creativecommons.org/licenses/by-nc/4.0/.
Statistics from Altmetric.com
If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.