Interpretable multiclassification machine learning models in predicting the extent of small cell lung cancer metastasis

Small cell lung cancer (SCLC) is known for its rapid disease progression and high metastasis, significantly reducing patient survival. The current lack of clinical models that can accurately predict the metastatic stage of SCLC limits the timeliness and effectiveness of treatment strategies. Machine learning techniques offer the possibility to improve the accuracy and operationalisation of predictions by utilising complex algorithmic pattern recognition. This study aimed to develop and interpret a predictive model for SCLC metastatic staging using real-world data. The Surveillance, Epidemiology, and End Results database (SEER, 2004-2017) of the National Cancer Institute was selected as the data source for this study, and multiple machine learning algorithms were used to predict the metastatic stage of SCLC. Lasso regression was used for feature selection and model parameters were optimised by cross-validation. Model performance was assessed by area under the receiver operating characteristic curve (AUC) and other statistical metrics. To enhance the interpretability of the model, SHAP (SHapley Additive exPlanations) was used to analyse the contribution of key predictor variables to model decisions. After comparing different machine learning models, the XGBoost model demonstrated the highest accuracy in predicting SCLC metastatic stages, especially in advanced metastatic stages (M1b and M1c), with significantly better AUC values than the other models. SHAP analyses revealed that age, tumour size, and treatment regimen were the main variables influencing the model predictions. LIME analyses further confirmed these variables' specific influence in the prediction of individual cases. This study successfully applied advanced machine learning methods in the prediction of SCLC metastatic stage, which significantly improved the accuracy of prediction and the interpretability of the model. These findings can help clinicians make more accurate treatment decisions in SCLC management, especially in strategy selection and early intervention.