Optimizing adjuvant therapy selection with machine learning after transoral surgery in HPV-related oropharyngeal
Andrea Costantino1, Vinidh Paleri2, Andrew Williamson2
1Otorhinolaryngology Unit, IRCCS Humanitas Research Hospital, Via Manni zo56, Milan, Rozzano 20089, Italy; Department of Otolaryngology - Head and Neck Surgery, AdventHealth Orlando, 410 Celebration Place, Celebration, FL 34747, United States.
Background:
De-escalation after transoral surgery (TOS) for HPV-related oropharyngeal squamous cell carcinoma (OPSCC) requires accurate risk stratification to identify patients who can safely omit adjuvant therapy without compromising survival. We evaluated whether machine-learning (ML) survival models improve postoperative risk grouping and guide adjuvant treatment selection.
Methods:
Using the National Cancer Database cohort of HPV-positive OPSCC treated with TOS (N = 5,569), we trained three survival ML models (DeepSurv, Neural Multi-Task Logistic Regression [NMTLR], Random Survival Forest [RSF]) on the surgery-alone subset (80:20 split). Each model generated cohort-wide risk scores that were mapped to three risk groups (low/intermediate/high). Within strata, multivariable Cox models estimated the association between adjuvant radiotherapy (RT) or chemoradiotherapy (CRT) and overall survival (OS).
Results:
All ML models showed good discrimination (C-index 0.77-0.80) and acceptable calibration. Risk stratification yielded clearly separated OS curves across low, intermediate, and high-risk. In DeepSurv-defined strata, adjuvant therapy conferred no measurable OS benefit in low-risk, benefit with RT or CRT in intermediate-risk, and the largest benefit with CRT in high-risk. NMTLR and RSF produced concordant treatment patterns, with model-specific variation in effect sizes. Feature-importance analyses consistently emphasized age, margin status, comorbidity, insurance status, and facility volume. The influence of extranodal extension varied by model.
Conclusions:
ML-based risk stratification after TOS refines adjuvant decision-making by aligning de-escalation (observation) to truly low-risk patients and escalation (RT/CRT) to those most likely to benefit. Prospective model-guided studies are warranted to assess the safety of RT omission in low-risk patients and the benefit of CRT over RT in high-risk.
