Impact of smoking on lung cancer patient survival: DT radiomics for smoking status prediction and causal survival analysis
File(s)
Accepted version
Supporting information
Author(s)
Type
Journal Article
Abstract
Background:
This study investigates the impact of smoking status (ever-smoker vs. never-smoker) on lung cancer survival, with smoking status considered an important prognostic factor in patients with lung cancer. We developed a CT-based radiomic model to predict smoking status using imaging features. The influence of smoking status on survival outcomes was further evaluated using the Cox Proportional Hazards (CPH) model. In addition, a Bayesian mediation analysis was performed to investigate the association between smoking status and survival and to explore potential pathways through which smoking status may be associated with patient outcomes.
Methods:
Retrospective CT scans from 2,062 patients with non-small cell lung cancer (NSCLC) were analyzed. An nnU-Net model was trained on the TCIA-NSCLC dataset for lung segmentation, validated using the OCTAPUS-AI dataset, and subsequently applied to the LIBRA dataset. Both datasets are multicenter studies conducted across the UK.
Radiomic features were extracted from the segmented lungs to develop a smoking status classification model, which was evaluated using both internal and external test sets. A Cox Proportional Hazards (CPH) model was then developed to investigate the effect of smoking status on survival prediction. Furthermore, to examine the pathways through which smoking status may influence survival outcomes, a Bayesian mediation analysis was incorporated into the Cox regression framework to evaluate the potential mediating effects of smoking-related factors.
Results:
The nnU-Net model accurately segmented lungs in the OCTAPUS-AI dataset and most LIBRA cases, although 18% of LIBRA scans with heterogeneous image intensities required refinement using an alternative method. The CT-based smoking status prediction model achieved an AUC of 0.779 (95% CI: 0.746–0.811) in the training cohort and 0.782 (95% CI: 0.718–0.841) in the independent test cohort for distinguishing ever-smokers from never-smokers. Incorporating smoking status significantly improved the discrimination of the overall survival model, increasing the C-index from 0.610 (baseline model) to 0.662 (model incorporating smoking status) (ΔC-index = 0.052; P = 0.006).. Bayesian mediation analysis estimated that only a small proportion of the association between patient-reported smoking status and survival was mediated by tumor stage.
Conclusions:
Smoking status can be estimated using CT-derived radiomics, and incorporating this information enhances survival prediction. Bayesian mediation analysis suggests that the association between smoking status and survival is driven predominantly by direct associations, with tumor stage contributing only a limited mediating role.
This study investigates the impact of smoking status (ever-smoker vs. never-smoker) on lung cancer survival, with smoking status considered an important prognostic factor in patients with lung cancer. We developed a CT-based radiomic model to predict smoking status using imaging features. The influence of smoking status on survival outcomes was further evaluated using the Cox Proportional Hazards (CPH) model. In addition, a Bayesian mediation analysis was performed to investigate the association between smoking status and survival and to explore potential pathways through which smoking status may be associated with patient outcomes.
Methods:
Retrospective CT scans from 2,062 patients with non-small cell lung cancer (NSCLC) were analyzed. An nnU-Net model was trained on the TCIA-NSCLC dataset for lung segmentation, validated using the OCTAPUS-AI dataset, and subsequently applied to the LIBRA dataset. Both datasets are multicenter studies conducted across the UK.
Radiomic features were extracted from the segmented lungs to develop a smoking status classification model, which was evaluated using both internal and external test sets. A Cox Proportional Hazards (CPH) model was then developed to investigate the effect of smoking status on survival prediction. Furthermore, to examine the pathways through which smoking status may influence survival outcomes, a Bayesian mediation analysis was incorporated into the Cox regression framework to evaluate the potential mediating effects of smoking-related factors.
Results:
The nnU-Net model accurately segmented lungs in the OCTAPUS-AI dataset and most LIBRA cases, although 18% of LIBRA scans with heterogeneous image intensities required refinement using an alternative method. The CT-based smoking status prediction model achieved an AUC of 0.779 (95% CI: 0.746–0.811) in the training cohort and 0.782 (95% CI: 0.718–0.841) in the independent test cohort for distinguishing ever-smokers from never-smokers. Incorporating smoking status significantly improved the discrimination of the overall survival model, increasing the C-index from 0.610 (baseline model) to 0.662 (model incorporating smoking status) (ΔC-index = 0.052; P = 0.006).. Bayesian mediation analysis estimated that only a small proportion of the association between patient-reported smoking status and survival was mediated by tumor stage.
Conclusions:
Smoking status can be estimated using CT-derived radiomics, and incorporating this information enhances survival prediction. Bayesian mediation analysis suggests that the association between smoking status and survival is driven predominantly by direct associations, with tumor stage contributing only a limited mediating role.
Date Acceptance
2026-08-26
Citation
Translational Lung Cancer Research
ISSN
2218-6751
Publisher
AME Publishing
Journal / Book Title
Translational Lung Cancer Research
Copyright Statement
Copyright This paper is embargoed until publication. Once published the author’s accepted manuscript will be made available under a CC-BY License in accordance with Imperial’s Research Publications Open Access policy (www.imperial.ac.uk/oa-policy).
License URL
Publication Status
Accepted