ESTRO 2025 - Abstract Book

S3766

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2025

Results: Among the 573 patients, four survival classes were defined: Class 0 (≤ 3 weeks), Class 1 (3–24 weeks), Class 2 (24–52 weeks), and Class 3 (≥ 52 weeks). The DT models selected eight predictive features: interleukin-8 (IL8), lymphocyte count, hemoglobin, Karnofsky index, steroid use, liver metastases, food intake variation, and biological equivalent dose. IL8, lymphocyte count, and hemoglobin were identified as the top predictors of survival ( Figure 1 ). The Gini based DT model performed better than the entropy-based model, especially in predicting short-term (≤ 3 weeks) and long-term (≥ 52 weeks) survival, with AUCs of 0.80 and 0.78, respectively. The model interpretability was enhanced through the hierarchical tree structure, allowing clinicians to understand the rationale behind survival predictions ( Figure 2 ). Figure 1 : feature importance ranking rercentage of the decision tree algorithms: (aO entropy-based model and (b) Gini impurity-based model.