Interpretable Machine Learning for Preliminary Lung Cancer Risk Assessment Based on Clinical Indicators and Lifestyle Factors
Open Access
Abstract
Background/Objective: Lung cancer remains a leading cause of cancer-related mortality, and access to low-dose CT screening is limited in resource-constrained settings. This study aimed to develop and evaluate an interpretable machine learning model for preliminary, symptom-based lung cancer risk discrimination using clinical indicators and lifestyle factors, without requiring imaging. Methods: A cross-sectional dataset of 561 participants (38 with a prior lung cancer diagnosis), following exclusion of a minors-eligible age bracket for which independent age verification was not possible, was analyzed. Five machine learning algorithms (Logistic Regression, Random Forest, XGBoost, LightGBM, CatBoost) were compared using a leakage-free pipeline, with ADASYN class balancing and hyperparameter tuning performed exclusively within cross-validation folds; feature selection was verified to be stable across folds (98% average overlap with the final feature set). The final model was selected solely on the basis of repeated cross-validated performance, without reference to the test set. Model interpretability was assessed using LIME, and a four-tier risk stratification system was developed and calibrated. Results: Random Forest achieved the best cross-validated F1-score (0.850 ± 0.092) among the five candidates and, given a clinically motivated preference for higher sensitivity at comparable F1, was selected as the final model. On the independent test set (n = 113), it achieved Accuracy = 0.956, Recall = 0.875, F1-score = 0.737, and AUC-ROC = 0.912. Sensitivity analysis indicated moderate reliance on a potential diagnostic proxy feature, and balancing-method comparisons (ADASYN, random oversampling, class weighting) confirmed that conclusions were not contingent on the specific resampling technique. Conclusions: The proposed model demonstrates solid discriminative performance and interpretability; however, given the cross-sectional design and limited number of positive cases, results should be interpreted as preliminary risk discrimination rather than validated early-detection prediction, warranting further prospective validation.Keywords
Funding Information
- Science Committee of the Ministry of Science and Higher Education of the Republic of Kazakhstan (AP23485656)
This publication has 22 references indexed in Scilit:
- Global burden of lung cancer in 2022 and projections to 2050: Incidence and mortality estimates from GLOBOCANCancer Epidemiology, 2024
- Predicting the risk of lung cancer using machine learning: A large study based on UK BiobankMedicine, 2024
- Application of Machine Learning Methods for Controlling Catalytic Cracking Plant2023
- Observed and relative survival trends of lung cancer: A systematic review of population‐based cancer registration dataThoracic Cancer, 2023
- OWL: an optimized and independently validated machine learning prediction model for lung cancer screening based on the UK Biobank, PLCO, and NLST populationsEBioMedicine, 2023
- Machine learning techniques for pulmonary nodule computer-aided diagnosis using CT images: A systematic reviewBiomedical Signal Processing and Control, 2022
- A Survey on Explainable Artificial Intelligence (XAI): Toward Medical XAIIEEE Transactions on Neural Networks and Learning Systems, 2020
- End-to-end lung cancer screening with three-dimensional deep learning on low-dose chest computed tomographyNature Medicine, 2019
- "Why Should I Trust You?"2016
- ADASYN: Adaptive synthetic sampling approach for imbalanced learning2008