Clinically Interpretable Survival Risk Stratification in Head and Neck Cancer Using Bayesian Networks and Markov Blankets.

Abstract

Purpose: To identify a clinically interpretable subset of survival-relevant features in head and neck (H&N) cancer using Bayesian Network (BN) and evaluate its prognostic and causal utility.

Methods and materials: We used the RADCURE dataset, consisting of 3,346 patients with H&N cancer treated with definitive (chemo)radiotherapy. A probabilistic BN was constructed to model dependencies among clinical, anatomical, and treatment variables. The Markov Blanket (MB) of two-year survival (SVy2) was extracted and used to train a logistic regression model. After excluding incomplete cases, a temporal split yielded a train/test (2,174/820) dataset using 2007 as the cutoff year. Model performance was assessed using area under the receiver operating characteristic (ROC) curve (AUC), concordance index (C-index), and Kaplan-Meier (KM) survival stratification. Model fit was further evaluated using a log-likelihood ratio (LLR) test. Causal inference was performed using do-calculus interventions on MB variables.

Results: The MB of SVy2 included 6 clinically relevant features: Eastern Cooperative Oncology Group (ECOG) performance status, T-stage, HPV status, disease site, the primary gross tumor volume (GTVp), and treatment modality. The model achieved an AUC of 0.65 and C-index of 0.78 on the test dataset, significantly stratifying patients into high- and low-risk groups (log-rank p < 0.01). Model fit was further supported by a log-likelihood ratio of 70.32 (p < 0.01). Subgroup analyses revealed strong performance in HPV-negative (AUC = 0.69, C-index = 0.76), T4 (AUC = 0.69, C-index = 0.80), and large-GTV (AUC = 0.67, C-index = 0.75) cohorts, each showing significant KM separation. Causal analysis further supported the positive survival impact of ECOG 0, HPV-positive status, and chemoradiation.

Conclusions: A compact, MB-derived BN model can robustly stratify survival risk in H&N cancer. The model's structure enables explainable prognostication and supports individualized decision-making across key clinical subgroups.