DE-ESCALATING INTENSITY AND PRESERVING OUTCOMES: A BAYESIAN-ML NETWORK META-ANALYSIS OF MULTIMODAL TREATMENT STRATEGIES IN HPV-POSITIVE OROPHARYNGEAL CANCER

This study presents the first integrative Bayesian-network meta-analysis enhanced by machine learning algorithms to evaluate and rank de-escalation treatment strategies for HPV-positive oropharyngeal squamous cell carcinoma (OPSCC). A total of 68,161 patients from 26 multicenter studies were included, encompassing randomized controlled trials, observational studies, and phase II investigations. Twelve distinct treatment strategies were analyzed, including TORS with de-escalated adjuvant RT, reduced-dose chemoradiotherapy, adaptive radiotherapy, and immunotherapy-based regimens. The SUCRA (Surface Under the Cumulative Ranking) scores revealed TORS + de-escalated RT as the top-ranked strategy (SUCRA = 0.91), followed by reduced-dose CRT (0.88) and adaptive RT (0.84). SHAP (SHapley Additive exPlanations) analysis from a Random Forest classifier confirmed that toxicity reduction (impact = 0.34) and QOL improvement (0.28) were the most critical features driving high SUCRA rankings, with overall survival (OS) rates consistently above 90% in the top three strategies. Funnel plots indicated low publication bias, and cluster heatmaps showed clear stratification of treatment profiles. The t-SNE visualization validated strong feature convergence among top-performing modalities. Our analysis demonstrates that machine learning-guided evidence synthesis can significantly enhance clinical decision-making in personalized OPSCC therapy by balancing oncologic efficacy with functional outcomes