Impact of consensus-based feature selection on machine learning performance for metastatic lung cancer prediction in resource-limited settings.

e20519 Background: In low- and middle-income countries (LMICs), lung cancer is frequently diagnosed at advanced or metastatic stages due to limited early detection infrastructure. Conventional logistic regression provides interpretability but limited predictive performance in high-dimensional settings, whereas penalized regression and machine-learning models can capture complex, non-linear relationships. Comparative evaluation of these approaches in LMIC metastatic cohorts remains scarce. We assessed regression-based and machine-learning models to predict metastatic lung cancer risk and identify key determinants, emphasizing the trade-off between predictive accuracy and clinical interpretability. Methods: We analyzed 4,480 lung cancer patients treated at the National Institute of Cancer Research and Hospital, Bangladesh (2021–2023) using 29 demographic, clinical, symptom-related, and laboratory variables. Missing binary and categorical data were imputed using observed-proportion sampling. Models evaluated included conventional logistic regression, LASSO-penalized logistic regression, ridge regression, Random Forest, and XGBoost. Feature importance was derived for each model and aggregated into a consensus ranking. The top 15 consensus-ranked features were used to re-train all models, with performance compared to full-feature models. Evaluation metrics included ROC–AUC, F1-score, sensitivity (metastatic recall), specificity, calibration plots, and decision curve analysis. Results: Among the top 15 consensus-ranked features, habitual factors emerged as the strongest predictors of metastatic disease, followed by poor performance status (ECOG ≥2), respiratory symptoms including breathlessness with chest pain, and multiple concurrent symptoms. After consensus-based feature selection, tree-based models demonstrated notable performance gains. Random Forest achieved a 3.6% increase in F1-score and 1.8% improvement in metastatic recall. XGBoost showed the largest relative gains: metastatic recall increased by 15.8%, F1-score by 8.3%, with modest improvements in discrimination. In contrast, regression models showed declines in F1-score (LASSO −7.9%, ridge −41.4%) despite stable or slightly improved specificity. Consensus-ranked features preferentially enhanced clinically relevant detection in tree-based models, demonstrating their superiority over regression approaches in this LMIC real-world dataset. Conclusions: Tree-based models achieved superior predictive accuracy and balanced sensitivity–specificity compared with conventional regression. Consensus-ranked features improved metastatic case detection despite dataset limitations inherent to LMIC settings. External validation is required to confirm generalizability in LMIC, real-world clinical environments.