Multicenter Radiomic Feature-Based Prediction of PD-L1 Status in Lung Cancer Using Multimodal PET/CT

With the growing utilization of immune checkpoint inhibitors (ICIs), there is an urgent need to identify biomarkers that can stratify patients into responders and non-responders based on programmed death-ligand (PD-L1) expression. The aim of this study is to determine the feasibility of building imaging-based predictive biomarkers for PD-L1 and system-atically evaluating a combination of several machine learning algorithms. A retrospective, multicenter study was conducted involving 467 advanced NSCLC patients eligible for ICIs across four various centers (France, China, Greece and Switzerland). Radiomic features derived from pretreatment PET/CT scans were used to construct predictive models for PD-L1 expression. After splitting the data 70/30 in training and testing sets, models to predict PD-L1 status were built using multivariate regression (MR) with 10-fold cross validation after feature selection based on LASSO, Random Forest (RF) and Support Vector Machine (SVM) both with embedded feature selection. All three pipelines used either the untransformed or harmonized features and were compared using Area Under the Curves (AUC). Among all 3 models, ComBat versions (whether pooled together of split by centers) out-performed untransformed (AUC 0.76 vs. 0.70, in MR with Lasso, AUC 0.78 vs. 0.73, in RF, and AUC 0.75 vs. 0.72, in SVM). This results explain how radiomics features can be used to train machine learning algorithms to predict PD-L1 expression status. Such models should be considered in future investigations to build robust and clinically relevant pre-dictive models.