Machine learning guided adjuvant treatment of head and neck cancer.

Abstract

6567 Background: A combined analysis of the EORTC 22931 and RTOG 95-01 trials confirmed that patients (pts) with head and neck squamous cell carcinoma (HNSCC) and positive margins or extracapsular extension (ECE) benefit from adjuvant chemoradiotherapy (CRT), but the best treatment of pts with other risk factors is unclear. We hypothesized that deep learning models could identify the margin/ECE negative pts who benefit from CRT. Methods: We abstracted pts from the NCDB diagnosed from 2004-2016 with resected HNSCC who received radiotherapy (RT). We reserved 20% of pts for validation and used the remaining 80% for feature selection and model training. Features were chosen based on independent significance in a Cox proportional hazards model, and included demographics, tumor stage, site, grade, RT dose, and receipt of chemotherapy. HPV status was included, and imputed when unknown. We generated survival predictions with DeepSurv (DS), random survival forest (RSF), and neural network multitask (NNM) models. We consider CRT to be recommended by a model if predicted survival is longer with CRT than RT. We calculated the median overall survival (mOS) difference and hazard ratio (HR) for receipt of treatment in line with model recommendations. This was repeated with inverse probability of treatment weighting (IPTW) to account for confounding. As a comparator, we used the intermediate risk factors in the EORTC (T3-4 except T3N0 larynx, N2-3, LVI, deep nodes with oral / oropharynx cancer) and RTOG (2 involved nodes) trials as decision rules. Results: 36,831 pts from the NCDB met the inclusion criteria. 92% had T3-4 or node positive disease, and 40% received CRT. RTOG, EORTC, DS, NNM, and RSF models recommend CRT for 32%, 74%, 63%, 61%, and 35% of pts. The concordance index in the validation set was 0.696, 0.692, and 0.699 for DS, NNM, and RSF. Treatment according to model recommendations in the validation cohort was associated with a mOS benefit of 18.4 months (7.6 to 29.3, 95% CI) for DS, 20.5 months (8.8 to 32.2, 95% CI) for NNM, and 5.8 months (-6.6 to 18.3, 95% CI) for RSF. Similar results were seen with IPTW. Conclusions: Machine learning models can predict benefit from CRT in margin/ECE negative pts, and outperform treatment according to EORTC or RTOG inclusion criteria in this cohort. External validation of these models is warranted. [Table: see text]