Evaluating Clinical Features and Prognostic Factors Associated With Response to Immune Checkpoint Blockade in mNSCLC Patients

AACR Annual Meeting 2021 Presentation
Authors Prerna Jain, Michelle Stein, Yinjie Gao, Denise Lau, and Aly A. Khan

In the past decade, immune checkpoint blockade (ICB) has become standard of care for patients with metastatic non-small cell lung cancer (mNSCLC). Biomarkers of response to ICB, such as PD-L1 status, help identify suitable candidates for ICB. However, differences in the underlying clinical features of patients and the association of such features with ICB outcomes has not received significant attention. We sought to evaluate clinical features in mNSCLC patients and determine their prognostic value in predicting response to ICB.

We retrospectively analyzed 1,903 patients with mNSCLC using real-world clinical data from the Tempus platform and generated a training and held-out test set with 1,351 and 552 patients, respectively. Clinically relevant features such as age, sex, comorbidity information, smoking status, and self-reported race/ethnicity were selected as clinical features and potential prognostic factors. We used a survival-SVM algorithm to model progression risk based on time to progression (TTP) from the start of first-line, FDA-approved treatment. The predicted risk output of the survival-model was used as a prognostic score for new or held-out patients. We evaluated the prognostic score on an independent clinical-only cohort from ASCO CancerLinQ (n = 3,217). We also examined the utility of the prognostic score by combining it with established biomarkers, such as PD-L1 status , using held-out clinicogenomic data from ICB-treated patients (n = 513). To evaluate model robustness, we examined performance on our held-out test set.

We found the combined use of all clinical features produced a more accurate prognostic score for predicting TTP (hazard ratio [HR] = 5.78, P = 1.20e-5) compared to individual clinical features (max individual HR from feature inputs = 1.22, P = 2.55e-3). Next, to evaluate the accuracy of the prognostic score more broadly in mNSCLC irrespective of treatment type, we analyzed an independent cohort of mNSCLC patients from ASCO CancerLinQ. We defined the top 15th percentile as high-risk for progression. We observed a significant stratification between high-risk and other patients (log-rank statistic = 101.68, P = 6.53e-24). Finally, we evaluated the model in the context of predicting ICB outcomes in mNSCLC. We found ICB-treated patients stratified based on prognostic score (15th percentile separation; log-rank statistic = 4.73, P = 0.029). However, when combined with PD-L1 status, we found better stratification of patients (15th percentile separation and PD-L1 positive; log-rank statistic = 14.653, P =1.29e-4) relative to PD-L1 status alone (PD-L1 positive; log-rank statistic = 9.671, P = 1.87e-3).

These results suggest that clinical and prognostic factors may augment established biomarkers and improve prediction of response to immune checkpoint blockade in mNSCLC patients.