Machine Learning-Driven Risk Stratification and Adjuvant Treatment Guidance in Oral Cavity Cancer.

Abstract

To develop and validate machine learning (ML) models for postoperative risk stratification in oral cavity squamous cell carcinoma (OCSCC) and to examine whether ML-derived risk groups modify the association between adjuvant therapy and overall survival (OS). Using the National Cancer Database, we identified adults with invasive OCSCC treated with primary surgery. The surgery-alone cohort (n = 18,543) was split 70/30 for training/testing to develop DeepSurv, Neural Multi-Task Logistic Regression (NMTLR), and Random Survival Forest (RSF) models. Risk scores were generated for the full cohort (n = 35,625) and converted to low, intermediate, and high groups. Within groups, treatment effects of adjuvant radiotherapy (RT) and chemoradiotherapy (CRT) were estimated using multivariable Cox models. The best performance was achieved by DeepSurv (C-index 0.73), with similar discrimination for NMTLR/RSF (C-index 0.71-0.72). For DeepSurv, the full cohort was partitioned into low- (50.0%), intermediate- (32.0%), and high-risk (18.0%) groups with distinct 5-year OS rates: 77.6%, 53.0%, and 29.3%, respectively. In the low-risk group, adjuvant RT (adjusted hazard ratios [aHR], 0.94 [95% CI, 0.87 to 1.02]) and CRT (aHR, 1.03 [95% CI, 0.91 to 1.17]) did not improve OS. In the intermediate-risk group, OS improved with RT (aHR, 0.61 [95% CI, 0.57 to 0.65]) and CRT (aHR, 0.56 [95% CI, 0.52 to 0.61]). In the high-risk group, both adjuvant RT (aHR, 0.47 [95% CI, 0.43 to 0.51]) and CRT (aHR, 0.39 [95% CI, 0.36 to 0.41]) were associated with improved OS compared with surgery alone. CRT was associated with a modest benefit compared with RT. NMTLR and RSF yielded concordant patterns. Top features included pT4a stage, age ≥70 years, and extranodal extension. ML-derived risk stratification identifies patients with OCSCC most likely to benefit from adjuvant therapy, supporting intensification for intermediate-/high-risk patients and potential deintensification for low-risk patients. External prospective validation is warranted to enable clinical implementation.