Quantifying Domain-Specific Risk Signals in Lung Cancer Severity Prediction: A Multi-Domain Ablation Study Using XGBoost and SHAP

Predictive modeling for lung cancer severity often struggles with the high dimensionality and multi-domain nature of risk factors. While individual contributors like smoking are well-documented, the relative predictive weight of lifestyle, environmental, and genetic domains remains insufficiently quantified in integrated frameworks. This study proposes an explainable machine learning approach using an XGBoost classifier to evaluate these three distinct risk domains. Utilizing the UCI Machine Learning Repository Lung Cancer Dataset, we implemented a domain-wise ablation study to isolate the predictive signal of each factor group. To ensure scientific rigor and address the “black box” nature of ensemble models, we employed 5-fold stratified cross-validation and SHAP (Shapley Additive Explanations) for feature-level transparency. Our results demonstrate that the integrated model achieves a classification accuracy of 95.7% (AUC-ROC = 0.98) on this dataset. Notably, ablation analysis revealed that the Lifestyle domain retained the highest standalone predictive performance (92.9%), followed by the Genetic/Clinical domain (94.6%), while the Environmental domain showed a more pronounced performance drop (73.3%), suggesting differential information density across risk categories. SHAP analysis identified cumulative smoking exposure as the primary feature influencing model predictions within this dataset. This study presents a proof-of-concept interpretable framework for lung cancer risk stratification, demonstrating that domain-wise ablation combined with explainable AI can provide transparent, feature-level insight to support rather than replace clinical judgment in settings where comprehensive diagnostic testing may be limited.