Automated three-dimensional radiomic body composition analysis enhances survival prediction in resectable non‑small cell lung cancer.

OBJECTIVE

To develop and validate a machine learning radiomics model integrating automated three-dimensional body composition and tumor imaging features for predicting overall survival in resectable non-small cell lung cancer (NSCLC).

MATERIALS AND

METHODS

This multicenter retrospective study included patients with resectable NSCLC treated between January 2013 and December 2017, who were assigned to training, internal, and external validation cohorts. A fully automated deep learning algorithm was developed for body composition segmentation. Radiomic features from tumor and body composition were extracted and integrated using extreme gradient boosting. Model performance was assessed using the concordance index (C-index) and time-dependent area under the curve (AUC), with interpretability evaluated by SHapley Additive exPlanations (SHAP). Kaplan-Meier analysis was performed for survival stratification.

RESULTS

Among 1,038 patients (mean age, 61.8 ± 10.7 years; 58.66% male), 293 (28.2%) died over a median follow-up of 3.31 years. In the training cohort, both tumor score and body composition score were independently associated with overall survival (hazard ratio 2.72 and 2.03, respectively; all p < 0.001). Incorporating body composition radiomics significantly improved discrimination compared with tumor-only models across cohorts (all p < 0.05). The comprehensive model, integrating clinicopathological factors, tumor score, and body composition score, demonstrated strong predictive capability for 1, 2, 3, and 5-year survival (AUCs > 0.80). SHAP analysis identified tumor score and body composition score as dominant predictors, stratifying patients into four phenotypes with distinct prognoses (all log-rank p < 0.05).

CONCLUSION

Integrating automated three-dimensional body composition with tumor radiomics enhances survival prediction and provides incremental value for postoperative risk stratification in resectable NSCLC.

KEY POINTS: Question Standard staging for resectable non-small cell lung cancer lacks objective three-dimensional quantification of host body composition, thereby limiting individualized prognostic assessment. Findings Machine learning models integrating automated three-dimensional body composition and tumor radiomics significantly outperform conventional tumor imaging in overall survival prediction. Relevance statement Automated three-dimensional body composition radiomics integrated with tumor imaging improves survival prediction in resectable non-small cell lung cancer, enabling more precise postoperative risk stratification and supporting individualized follow-up and supportive care strategies.