Benchmarking of radiobiological NTCP models in head and neck radiotherapy using independent computational pipelines: an institutional validation study with machine learning augmentation.

Abstract

Background & purpose: Normal tissue complication probability (NTCP) models require institutional validation before clinical implementation. Traditional radiobiological models, such as the Lyman-Kutcher-Burman (LKB) and Equivalent Uniform Dose (EUD) models, provide mechanistic dose-response frameworks, while machine learning (ML) approaches offer exploratory, data-driven alternatives that remain inadequately characterised in South Asian populations.

Methods: This retrospective study included 51 head and neck cancer patients treated with definitive radiotherapy. Binary endpoints were Grade ≥2 xerostomia (n = 3), dysphagia (n = 5) and mucositis (n = 4), scored using Common Terminology Criteria for Adverse Events version 5.0. NTCP calculations were performed using two independent computational pipelines (MATLAB-based RBMODELv1 and a Python implementation), with agreement assessed using Bland-Altman analysis. Traditional NTCP models (LKB, EUD) were evaluated and compared with artificial neural networks and XGBoost in a hypothesis-generating framework using a stratified 70:30 train-test split. Model performance was assessed using the area under the receiver operating characteristic curve (area under the curve), accuracy and Spearman's rank correlation.

Results: Excellent agreement was observed between computational pipelines (mean bias 0.8%, 95% limits -1.9% to 3.5%). Traditional models demonstrated strong rank-order correlation with toxicity grades (ρ = 0.61-0.79, p < 0.001) and high accuracy (LKB: 90.0%-94.1%). Institution-specific parameters differed from quantitative analyses of normal tissue effects in the clinic values, including a lower parotid TD50 (34.1 versus 39.0 Gy). Exploratory ML analyses showed numerically higher discrimination for parallel organs but not for mixed-architecture structures; however, severe class imbalance (3-5 events per endpoint) limits statistical reliability.

Conclusion: Dual computational pipelines enable reproducible NTCP modeling for institutional use. Traditional radiobiological models perform acceptably after local calibration, while exploratory ML findings suggest potential organ-architecture-dependent patterns that require validation in adequately powered multi-institutional cohorts.