Machine learning-based prediction of overall survival after lung SBRT using clinical, dosimetric, and radiomics features: a multicenter study

Radiomics combined with clinical and dosimetric data may improve outcome prediction after lung stereotactic body radiotherapy (SBRT). This study aimed to develop machine-learning models to predict overall survival (OS) after lung SBRT for primary and metastatic lung tumours using planning-CT-derived features. Clinical and dosimetric data were retrospectively collected from six centres. Radiomic features were extracted using double tumour segmentation and filtered for reproducibility (intraclass correlation ≥ 0.80). From 181 lesions in 158 patients, 180 features (124 radiomic, 45 clinical, 11 dosimetric) were aggregated to the patient level. OS was modelled as a time-to-event outcome using random survival forests, per tumour origin and pooled with tumour origin as a covariate; features were selected by LASSO within cross-validation folds and interpreted with SHAP. The pooled model achieved a concordance index of 0.62 (95% CI 0.54–0.71) with time-dependent AUC of 0.69, 0.62 and 0.52 at 1, 2 and 3 years; a temporal split preserved this (0.61) whereas leave-one-centre-out validation gave a pooled concordance index of 0.57 (95% CI 0.48–0.65). Adding dosimetric and radiomic features to a clinical model increased discrimination (ΔC + 0.12, 95% CI 0.001–0.244). Recurrence- and metastasis-free survival, examined as secondary endpoints, were near chance. These leakage-free, censoring-appropriate models provide an honest, hypothesis-generating benchmark for OS prediction after lung SBRT, most informative at one year, that will require external validation.