TY  - JOUR
T1  - Explainable Ensemble Machine Learning Framework for Lung Cancer Status Classification: A Retrospective Cross-sectional Study 
AU  - Saad, Hamza
JF  - Cancer Screening and Prevention
VL  - 
IS  - 000
SN  - 2835-3315
SP  - 
EP  - 
Y1  - 2026-09-30
DO  - 10.14218/CSP.2026.00009
UR  - https://www.xiahepublishing.com/2835-3315/CSP-2026-00009
AB  - Background and objective 
Machine-learning approaches that combine predictive performance with model interpretability may improve lung cancer status classification. This study aimed to develop and internally evaluate an Explainable Precision Screening Framework for lung cancer risk classification using demographic, behavioral, and symptom-based variables from a publicly available dataset.
Methods
This retrospective cross-sectional study analyzed a publicly available Kaggle dataset containing 309 records, including 270 labeled as lung cancer and 39 as non-cancer. Six models—logistic regression, support vector machine (SVM), random forest, LightGBM, XGBoost, and a stacking ensemble—were compared using a stratified hold-out test set and repeated stratified five-fold cross-validation. Performance was assessed using classification metrics with bootstrap 95% confidence intervals (CIs). A separate Shapley additive explanations (SHAP) analysis was applied to the standalone SVM model for exploratory feature attribution.
Results
Across all models, accuracy ranged from 85% to 92% (ROC-AUC: 0.93–0.95). The stacking ensemble achieved 0.92 accuracy (95% CI: 0.85–0.98), 0.94 sensitivity (95% CI: 0.88–1.00), 0.75 specificity (95% CI: 0.40–1.00), 0.96 precision (95% CI: 0.89–1.00), 0.95 F1 score (95% CI: 0.91–0.99), and 0.95 ROC-AUC (95% CI: 0.89–0.99). Its PR-AUC was 0.993 (95% CI: 0.981–0.999), and its Brier score was 0.074 (95% CI: 0.037–0.121). SHAP analysis identified smoking, yellow fingers, coughing, chest pain, wheezing, shortness of breath, and age as the features contributing most strongly to its predictions.
Conclusions
Within this public retrospective dataset, the stacking ensemble was among the highest-performing models, whereas separate SHAP analysis of the standalone SVM model identified the features contributing most strongly to its predictions. Given the marked class imbalance, the absence of a reported clinical reference standard for the source labels, the framework requires independent validation in clinically verified multicenter cohorts before clinical use.