A retrospective multicenter study has demonstrated that combining planning computed tomography radiomics with clinical and dosimetric data can modestly improve machine learning predictions of overall survival following lung stereotactic body radiotherapy. Researchers analyzed data across six medical centers to develop predictive models for patients treated for primary and metastatic lung tumours. By integrating quantitative imaging descriptors with radiation dosimetry and clinical records, the team sought to address the clinical challenge of prognosticating survival trajectories after focal ablative lung radiotherapy.
The underlying patient cohort comprised 158 individuals presenting with a cumulative total of 181 lung lesions. To extract robust quantitative image characteristics from baseline planning computed tomography scans, investigators implemented a double tumour segmentation workflow. Extracted radiomic candidate features were systematically screened for reproducibility, retaining only those variables meeting or exceeding an intraclass correlation coefficient threshold of 0.80. Following this quality filtering step, a final set of 180 variables was aggregated to the patient level, consisting of 124 radiomic features, 45 clinical variables, and 11 dosimetric parameters.
To handle survival analysis without introducing methodological bias, the researchers formulated overall survival as a time-to-event outcome. The modeling framework relied on random survival forests, which were constructed both separately by tumour origin and as a pooled model that incorporated tumour origin as an explicit covariate. Feature selection was conducted strictly using the least absolute shrinkage and selection operator (LASSO) embedded within internal cross-validation folds to avoid data leakage. Furthermore, the investigators applied SHapley Additive exPlanations (SHAP) to interpret the relative contributions and influence of selected features within the survival forests.
The resulting pooled random survival forest model achieved an overall concordance index of 0.62, with a 95 percent confidence interval ranging from 0.54 to 0.71. An evaluation of time-dependent discriminative capacity revealed that model accuracy diminished over follow-up duration. Specifically, the time-dependent area under the receiver operating characteristic curve was 0.69 at one year, dropping to 0.62 at two years, and declining to 0.52 at three years post-treatment. These temporal metrics indicate that the predictive utility of the baseline features was concentrated primarily in the immediate post-radiotherapy period.
The integration of multimodal data proved superior to relying on standard clinical indicators alone. Incorporating dosimetric and radiomic metrics into a baseline clinical model produced a statistically significant gain in discriminatory power, yielding an increase in concordance index of 0.12, with a 95 percent confidence interval spanning from 0.001 to 0.244. However, model robustness varied across validation schemes. While an internal temporal split preserved discrimination with a concordance index of 0.61, a leave-one-centre-out cross-validation across the participating institutions yielded a lower pooled concordance index of 0.57, accompanied by a 95 percent confidence interval between 0.48 and 0.65.
In contrast to the findings for overall survival, secondary oncological endpoints proved far more challenging to forecast. When the researchers tested the machine learning framework on recurrence-free survival and metastasis-free survival, prediction discrimination dropped to near-chance levels.
The study authors concluded that their methodological approach—characterized by leakage-free feature selection and censoring-appropriate survival modeling—provides an honest, hypothesis-generating benchmark for post-SBRT survival modeling. Given that the models demonstrated their highest discriminatory value at the one-year mark and exhibited sensitivity to center-level variations during leave-one-centre-out validation, the investigators emphasized that comprehensive external validation across broader independent cohorts will be essential before such tools can be considered for clinical implementation.
Subject of Research: Machine learning prediction of overall survival after lung stereotactic body radiotherapy
Article Title: Machine learning-based prediction of overall survival after lung SBRT using clinical, dosimetric, and radiomics features: a multicenter study
Article References: Invernizzi, C., Benveniste, P.-L., El Ayachy, R., Leroy, T., Blanchard, N., Schipman, B., Debelleix, C., Chamois, J., Pointreau, Y., & Bibault, J.-E. (2026). Machine learning-based prediction of overall survival after lung SBRT using clinical, dosimetric, and radiomics features: a multicenter study. Scientific Reports. https://doi.org/10.1038/s41598-026-77575-6
Image Credits: AI Generated
DOI: 10.1038/s41598-026-77575-6
Keywords: stereotactic body radiotherapy, lung SBRT, radiomics, machine learning, random survival forests, overall survival, planning CT, dosimetric features, LASSO feature selection, SHAP interpretation, concordance index, multicenter validation, leave-one-centre-out validation, tumour segmentation, intraclass correlation, time-dependent AUC
Cite Scienmag News
Teresa Odom. (November 17, 2026). Multicenter Machine Learning Study Predicts Overall Survival After Lung SBRT. Scienmag. https://scienmag.com/multicenter-machine-learning-study-predicts-overall-survival-after-lung-sbrt/
Teresa Odom. "Multicenter Machine Learning Study Predicts Overall Survival After Lung SBRT." Scienmag, 17 November 2026, https://scienmag.com/multicenter-machine-learning-study-predicts-overall-survival-after-lung-sbrt/. Accessed 17 November 2026.
Teresa Odom. "Multicenter Machine Learning Study Predicts Overall Survival After Lung SBRT." Scienmag. November 17, 2026. https://scienmag.com/multicenter-machine-learning-study-predicts-overall-survival-after-lung-sbrt/








