Correlation of clinical benefit from immune checkpoints and baseline radiomic signature in inoperable NSCLC without activating mutations.

Abstract

e20571 Background: Checkpoint inhibitors (ICI) are an important part of the NSCLC treatment algorithm. Existing criteria for ICI efficacy prognostication rely on exclusion of activating mutations carriers and PD-L1 expression. Still the sensitivity of this approach is not sufficient and significant percentage of patients receive ineffective expensive treatment with an elevated risk of side effects. We aimed to define whether additional information drawn from the baseline CT scans and interpreted with a machine learning model might help to improve the accuracy of selection for IO clinical benefit. Methods: We included patients with inoperable NSCLC treated in 1 st or 2 nd line with ICI (pembrolizumab and bioanalogues, atezolizumab, nivolumab or prolgolimab) either as monotherapy or a combination. Clinical benefit was defined as overall survival equal or above 24 months. Tumor volume was marked by a certified radiologist using 3D Slicer. Our radiomics pipeline integrated four types of normalization (none, MinMax, Z-score, Mean), preprocessing methods (ANOVA, RFE, KW, Relief), ten classifiers (SVM, LDA, Logistic Regression, AdaBoost, Gaussian Process, AE, Random Forest, LR-Lasso, Decision Tree, Naïve Bayes), and instruction to classifiers to select 2–20 radiomics features. This resulted in construction of 6,080 models using Feature Explorer (FAE) programmed with Python and utilizing NumPy, pandas and scikit-learn modules. Results: We included CT scans from 239 sequential patients with inoperable NSCLC. The best prognostically performing model was selected based on the AUC, accuracy, MCC, PPV and NPV in the validation datasets and independent internal test set. An optimal model was achieved by combining z-score normalization, the removal of outcome-unassociated features by ANOVA, and LASSO feature selection. It included only two out of 2141 calculated radiomics features: wavelet-LLL-NGTDM-strength and Logarithm-GLSZM-SizeZoneNonUniformityNormalized. Such small model improved the model’s ability to generalize to unseen data, achieving an AUC of0.792, accuracy of 84.7%, MCC of 0.533, PPV of 0.471 and NPV of 0.962 in the independent internal test set. The best-performing model (age, platelet count) including only clinicopathological features achieved a predictive AUC of 0.680 and an accuracy of 72.3%, while the best model combining both clinicopathological and radiomics features reached the AUC of 0.816 with an accuracy of 81.2%. Conclusions: Overall, our study highlights that radiomics can achieve very good performance in the prognostication of clinical benefit, defined as prolonged overall survival in patients with lung cancer treated with ICI. Prognostic models developed in this study may potentially provide a clinical predictor of disease outcome outperforming routinely used clinicopathological parameters.