ESTRO 2025 - Abstract Book

S3866

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2025

Results: FSSVM survival models achieved the best performance for the clinical-only prediction with a concordance index (C index) of 0.795 ± 0.028. The GBSA survival models yielded the best C-index of 0.729 ± 0.078 for the pathology text only model. The early-fusion model achieved the best performance of 0.749 ± 0.027 with the FSSVM model. The best models stratified patients into low, intermediate, and high-risk subgroups, with KM curves and log-rank tests showing significant OS differences: high vs. intermediate (p = 0.001), intermediate vs. low (p = 0.005), and high vs. low (p < 0.0001) (Figure 2).

Conclusion: A configurable, modular, end-to-end outcome prediction pipeline was developed using LLMs, ensemble and SVM based survival models. This approach accurately stratified HNC patients into clinically significant subgroups by using only pre-treatment information, such as tabular clinical data and unstructured pathology reports. It also has the potential to facilitate patient selection in ongoing clinical trials for treatment strategies in HNCs.