ESTRO 2025 - Abstract Book

S1033

Clinical – Head & neck

ESTRO 2025

2496

Poster Discussion Federated Prediction Models and External Validation for Radiotherapy Outcomes in Head and Neck Cancer using F.A.I.R. Clinical and Radiomics Data Varsha Gouthamchand 1 , Benedetta Gottardelli 2 , Gauri Kulkarni 3 , Umesh B Sherkhane 1 , Joshi Hogenboom 1 , Rajamenakshi Subramanian 3 , Ashish Kumar Jha 4 , Sneha Mithun 1 , Nilendu C Purandare 4 , Jai Prakash Agarwal 5 , Krithikaa Sekar 6 , Lohith G 6 , Sarbani G Laskar 5 , Shwetabh Sinha 5 , Frank JP Hoebers 1 , Venkatesh Rangarajan 4 , Gaur Sunder 3 , Andre Dekker 1 , Johan van Soest 1 , Leonard Wee 1 1 Department of Radiation Oncology (Maastro), Maastricht University, Maastricht, Netherlands. 2 Department of Diagnostica per Immagini, Radioterapia Oncologica ed Ematologia, Università Cattolica del Sacro Cuore, Rome, Italy. 3 The Centre for Development of Advanced Computing, C-DAC, Pune, India. 4 Department of Nuclear Medicine and Molecular Imaging, Tata Memorial Hospital, Mumbai, India. 5 Department of Radiation Oncology, Tata Memorial Hospital, Mumbai, India. 6 Healthcare Global Enterprises Ltd, HCG, Bangalore, India Purpose/Objective: Integrating data from multiple centers for predictive modeling in oncology poses significant challenges, particularly in ensuring model generalizability while maintaining data privacy. This study addresses these challenges in head and neck cancer (HNC) by utilizing datasets from institutions in India and the Netherlands as part of the TRAIN initiative [1]. Following the Findable-Accessible-Interoperable-Reusable (F.A.I.R.) principles [2], diverse data types - including clinical information, radiomics features, and Digital Imaging and Communications in Medicine (DICOM) metadata- were harmonized into a unified semantic framework using Linked Data. This integrated data was then used for federated Cox Proportional Hazards (CPH) modelling using the privacy-enhancing Personal Health Train (PHT) [3] method. Material/Methods: The open-source infrastructure, Vantage6 (v3.7.3) [4] was used to create a federated privacy-enhancing data exploration and data discovery dashboard. Quality checks, data harmonization and case-mix summarization were performed without any patient personal data transfer. Our distributed pipeline incorporated Correlation-Based Feature Selection (CFS) and LASSO-regularized Cox regression to identify optimal predictors for CPH models of Overall Survival (OS), Distant Metastasis (DM), and Locoregional Recurrence (LRR). All CPH models were developed using the same FAIR-ified datasets and same PHT infrastructure. Six distinct institutional datasets, both public and private, of 1131 oropharyngeal cancer patients were used in this study. We evaluated four models using three types of features: clinical-only (C-Model), radiomics-only from the primary tumor (RP-Model) and radiomics-only from malignant neck nodes (RN-Model), and finally a combined model (CR-Model) combining the linear predictors from the other three models. Results: The C-Model demonstrated the best predictive performance for OS and DM, achieving global average C-indices of 0.62 and 0.64, respectively, across external validation cycles. Radiomics-based models were superior for LRR, with both RN-Model and CR-model achieving a C-index of 0.64 in external validation. Conclusion: This study demonstrates the feasibility of federated multi-centric collaborations for predictive model development in oncology, leveraging F.A.I.R. principles and secure PHT infrastructure to integrate diverse data sources. These results suggest there is added value in including pre-treatment radiomics features in time-to-event prediction models. The TRAIN project demonstrates the potential of combining multi-institutional to improve model performance, offering a scalable framework for reaching personalized oncology care.

Keywords: F.A.I.R., Federated Learning, Predictive modelling