Background: We previously showed that in head and neck squamous cell carcinoma (HNSCC) treated with immune checkpoint inhibitors (ICIs), pretreatment higher lactate dehydrogenase (LDH) and absolute (abx) neutrophils as well as lower percent (%) lymphocytes correlated with worse overall survival (OS). Using these peripheral blood biomarkers (PBBMs), we developed a prognostic signature for OS risk stratification in HNSCC treated with ICIs. We then validated our signature in an independent cohort.
Methods: We randomly split our institutional dataset of recurrent/metastatic HNSCCs treated with ICIs (n=151) into training (n=100) and testing (n=51) datasets. Using the training data, we built an OS multivariable Cox regression model with LDH, % lymphocytes, and abx neutrophils and adjusted for ECOG, p16-status, and smoking as confounders. Combined proportion score (CPS) was not included as it had not been routinely collected. The continuous risk score = ß1*X_LDH + ß2*X_%lymphocytes + ß3*X_abxneutrophils was calculated based on estimated coefficients from the Cox model and 3 PBBMs. We trichotomized patients into risk groups using a grid search maximizing score test statistics from the Cox model to identify optimal cutpoints (θ1, θ2) defining low-risk (risk score≤θ1), intermediate-risk (θ1θ2). Kaplan-Meier OS curves were generated by risk group. We validated the continuous risk score and trichotomized signature in the testing data by Cox regression. Using an independent cohort as the external validation set (n=54), we validated the trichotomized signature by Cox regression.
Results: Training and testing datasets showed no significant differences in age, sex, race, smoking, alcohol, Charlson comorbidity index, p16-status, ECOG, or PBBMs. From the training data, the OS risk score = 1.2431*log10(LDH) - 1.9464*log10(%_lymphocytes) + 0.473*log10(abx_neutrophils). Optimal risk score cutpoints were θ1=0.401 and θ2=1.029, resulting in 21 low-risk, 44 intermediate-risk, and 35 high-risk patients, with Kaplan-Meier curves for the trichotomized signature shown in Figure 1A. Testing dataset risk scores were then calculated and trichotomized into risk groups, resulting in 7 low-risk, 24 intermediate-risk, and 20 high-risk patients; Kaplan-Meier curves showed clear separation of the high-risk group but some overlap between low and intermediate-risk groups (Figure 1B). Adjusting for ECOG, smoking, and p16-status in the testing group, higher risk scores correlated with worse OS (HR 2.08, [95%CI 1.17-3.68], p=0.012). The prognostic signature was then evaluated in an independent cohort. There were no significant clinicodemographic differences across the training, testing, and independent datasets, except that ICIs were more frequently first-line therapy in the independent dataset. Trichotomizing risk groups using the previously determined cutpoints was borderline significant (p=0.068), with Kaplan-Meier curves showing the low-risk group with best OS and some overlap between intermediate and high-risk groups (Figure 1C).
Conclusions: We developed a prognostic biomarker signature for OS based on previously identified PBBMs for HNSCC treated with ICIs. Validation of the prognostic signature in both the testing and independent cohorts were promising, although limited by sample size. This prognostic model may aid in identifying patients less likely to benefit from ICIs and warrants validation in a larger cohort receiving first-line ICIs as well as correlation with CPS biomarker.