Accurate lung cancer staging is crucial for prognosis and treatment planning. In this study, we present an autonomous multi-agent machine learning framework for multiclass staging of lung cancer by integrating multi-model machine learning, explainable AI, and computational complexity assessment to enhance interpretability and clinical relevance. The framework trains Random Forest, XGBoost, and CatBoost classifiers on a fully preprocessed dataset of N samples, which includes demographic, clinical, and genomic profiles of the cases. The best-performing model is automatically selected based on test accuracy. Class imbalance is addressed using SMOTE, and features are standardized to improve model generalization. The proposed framework achieved an overall accuracy of 98%, outperforming several recent lung cancer staging models that reported accuracies between 91 and 95% and AUC values ranging from 0.91 to 0.94. The framework also achieved near-perfect multiclass AUC-ROC values (0.997–1.000), demonstrating highly robust discrimination across all clinical stages. Explainable AI analysis using LIME revealed that features such as Metastatic site, Primary Tumor Site, Race Category, and TP53 pathway were influential in model predictions, aligning with biological and clinical understanding. Computational complexity analysis indicated linear scaling of training time with dataset size while testing remained highly efficient, providing real-time applicability. These results highlight the effectiveness of framework in accurate, interpretable and scalable multiclass lung cancer staging and offering a valuable tool for precision oncology.
Kayani et al. (Wed,) studied this question.