Development and Validation of a Machine Learning-Based Predictor for OS and PFS in HPV-Negative HNSCC Patients With Microscopic ENE and Intermediate-Risk Disease

PURPOSE/OBJECTIVE(S) Improved stratification is needed in HPV-negative HNSCC patients with intermediate and select high-risk disease. Significant molecular and immune response features have not been well-integrated with clinicopathologic factors to predict outcomes in this population. We sought to develop and validate an integrated molecular and clinicopathologic machine learning-based predictor for OS and PFS in this population to inform a future trial design. We hypothesize that our predictor will stratify patients better than a predictor derived from standard techniques. MATERIALS/METHODS We included 253 patients from TCGA with pathologic stage III-IVB (excluded T4b; included N3a) HPV-negative HNSCC with microscopic ENE and intermediate-risk disease (close margins/LVSI/PNI), treated with surgery and RT +/- chemotherapy. We split the data into training (70%) and testing (30%) sets. We identified 29 relevant molecular features (genomic: mutation vs. no mutation, transcriptomic: ≥ vs. < upper third quartile of log2 transformed mRNA expression) associated with significant HNSCC pathways, including cellular proliferation, cellular differentiation, cell cycle control, adhesion and invasion, anti-tumor immune response, and apoptosis. We also identified 19 demographic, behavioral, and clinicopathologic factors. We performed random survival forest modeling on OS and PFS and validated the machine learning-based predictor on the testing set, using ROC/AUC. Variables of importance were identified using the "Janitza" method. We also built a predictor using standard "best" Cox proportional hazards modeling and compared AUC values for OS and PFS versus our machine learning-based values. RESULTS The median OS and PFS times were 4.4 years and 3 years, respectively. After accounting for pairwise correlations, we kept 38 variables for model building. Using the training set, significant variables of most importance for OS included age, female sex, PNI, LVSI, ≥ four pathologically involved lymph nodes, microscopic ENE, TP53 missense and nonsense mutations, and expression of PD1, LAG3, TIM3, L1CAM, CASP8, PIK3CA, E2F2, E2F4, FAT1, and NOTCH1 (all P-values < .05). Significant variables of importance for PFS included age, female sex, ≥ four pathologically involved lymph nodes, microscopic ENE, anatomic subsite, TP53 nonsense mutations, and expression of PD1, L1CAM, PIK3CA, and FAT1 (all P-values < .05). Using the testing set, we validated the 38-variable predictor on OS and PFS with AUC values of 0.82 and 0.75, respectively. In contrast, the "best" Cox models resulted in low AUC values for OS and PFS (0.65 and 0.66, respectively). CONCLUSION We developed and validated a machine learning-based predictor for OS and PFS that outperforms standard Cox predictors in HPV-negative HNSCC. This study provides a rationale to use the predictor as a stratifier in a prospective clinical trial of surgery, radiotherapy, and PD1 blockade in patients with microscopic ENE and intermediate-risk disease.