Lymphocyte Kinetics is an Important Prognosticator in Predicting Survival Outcomes for Head and Neck Squamous Cell Carcinoma (HNSCC) Patients Using a Ransom Survival Forest (RSF) Model

Abstract

<h3>Purpose/Objective(s)</h3> Lymphopenia is associated with poor outcomes in HNSCC, yet there is no consensus on how we can use lymphocyte metrics (LMs) dynamically in personalized treatment. We hypothesize that treatment-related lymphopenia (TRL) has greater prognostic significance vs. baseline lymphopenia (BL) in predicting OS & PFS, and that the time-varying attributes of lymphocyte metrics (LMs) have greater prognostic significance vs. static LMs. Given the number of covariates necessary to test our hypothesis, we need to move beyond traditional Cox regression in favor of a robust, machine learning model that is completely non-parametric, allows interactions & leads to lower risks to overfitting. <h3>Materials/Methods</h3> Eligibility criteria were newly diagnosed M0 HNSCC patients treated with neck RT ± chemo in 2015-19 with baseline absolute lymphocyte counts (ALC) and ≥ 1 on-treatment ALC data point. We retrospectively analyzed prospectively acquired data on 23 covariates: 6 baseline, 5 disease, 6 chemoRT & 6 LMs (baseline ALC, min ALC, max ALC decrease, baseline neutrophil-to-lymphocyte ratio or NLR, max NLR & max NLR increase). Overall, there was 1.4% missing data. We used random survival forest (RSF), an ensemble-based method for overall & progression-free survival analysis (OS & PFS). RSF is attractive because it handles missing data by proximity imputation, is not affected by multicollinearity, and provides internal validation of model fit using 37% of out-of-bag (OOB) data that do not participate in model training. Averaging over 1000 survival trees, we divided the OOB ensemble predictor scores into tertile risk groups. Recognizing the burden to analyze 23 covariates, we developed a simplified RSF model with 10 binary covariates using conditional inference tree & weighted risk scores using ranked variable importance (VIMP). We also studied pairwise interactions amongst covariates. <h3>Results</h3> 316 patients were eligible for analysis. Both full & simplified models showed show good predictive performance with survival AUCs 0.71-0.76 for OS & 0.66 – 0.72 for PFS. Two-year OS for low, intermediate, and high-risk groups was 94 vs 89 vs 70% in the full model (FM) and 96 vs 80 vs 65% in the simplified model (SM). Similarly, 2-year PFS was 89 vs 72 vs 62% (FM), and 88 vs 71 vs 49% (SM). Out-of-bag (OOB) prediction errors were 0.26 – 0.34. Consistent in both models, these 4 covariates contributed the most to survival outcomes: age, KPS, RT dose and max ALC decrease. Max ALC decrease was ranked #4 in relative variable importance (VIMP) score, indicating a greater influence on OS & PFS than HPV, chemo type/dose, and other LMs (range VIMP rank 6-20). Of note, GTV was more influential vs. T&N stage (rank 5-7 vs 9-22). <h3>Conclusion</h3> Our data suggests that TRL has a greater prognostic significance when compared with BL. Max ALC decrease has superior predictive performance vs other LMs. To our knowledge, this is the first study that highlights the time varying attributes of ALC as an important prognosticator for HNSCC outcomes.