ESTRO 2025 - Abstract Book

S3366

Physics - Machine learning models and clinical applications

ESTRO 2025

Switzerland. 5 Department of Radiation Oncology, Centre Léon Bérard, Lyon, France. 6 Department of Radiation Oncology, Inselspital, Bern University Hospital, Bern, Switzerland. 7 Department of Pathology, Klinikum Stuttgart, Stuttgart, Germany. 8 Institute of Tissue medicine and Pathology, Inselspital, Bern University Hospital, Bern, Switzerland. 9 Department of Radiation Oncology, University Medical Center Groningen, Groningen, Netherlands Purpose/Objective: Head and neck squamous cell carcinomas (SCCs) often metastasize in the neck through lymphatic pathways, leading to elective irradiation of lymph node levels (LNLs) at risk for occult metastases. Current guidelines are based on overall lymph node involvement prevalence, lacking individualized patient-specific predictions. 1 We previously developed an interpretable probabilistic model predicting lymphatic spread in oropharyngeal SCCs, treating each LNL as a binary random variable (involved or healthy) based on learned spread probabilities. 2 This study extends that model to additional tumor sites including oral cavity and hypopharynx. The model predicts the risk of occult involvement in LNLs, incorporating detailed tumor subsite information using ICD codes, T-category, and clinical nodal status. Material/Methods: The training data includes seven datasets from 5 European centers with 1634 patients, each characterized by primary tumor subsite and pathological or radiological LNL status. To capture the continuous variation in spread patterns across subsites (Figure1), we apply a mixture model with three lymphatic spread components instead of separate models for each subsite. The mixture model assigns weights to each tumor subsite, indicating its similarity to each model-component.

Results: The mixture model assigns model-components to oral cavity-like tumors, oropharynx-like tumors and hypopharynx like tumors (Figure2). All subsites are described as weighted combinations of the components. For instance, palatine