Machine learning predictors of long-term survival in advanced lung cancer immunotherapy: A comparative analysis of classification approaches

BackgroundImmunotherapy has transformed outcomes for advanced non-small cell lung cancer (NSCLC), but substantial heterogeneity in treatment response necessitates personalized predictive approaches. The emergence of long-term survivors exhibiting plateau patterns in survival curves suggests distinct subpopulations requiring identification.MethodsWe conducted a retrospective analysis of 1801 patients with advanced NSCLC enrolled in 14 clinical trials (2002-2025) evaluating three immunotherapies: CIMAvax-EGF, Racotumumab, and Nimotuzumab. Two classification approaches were compared: mixture survival model and percentile-based stratification. Eight machine learning algorithms (logistic regression, decision tree, random forest, support vector machine, k-nearest neighbors, CatBoost, LightGBM, XGBoost) were trained and evaluated using accuracy, AUC-ROC, precision, and recall metrics.ResultsThe percentile-based approach (33rd and 66th percentiles) outperformed the mixture model in identifying long-term survivors, achieving balanced precision (0.60-0.6) and recall (0.60-0.77) across both groups. Logistic regression and support vector machine demonstrated the best performance with 67% accuracy and 0.72 AUC. Key predictors consistently identified across models included neutrophil count, body mass index, and a complete response to first-line chemotherapy.ConclusionsMachine learning approaches, particularly when combined with percentile-based survivor classification, effectively predict long-term responders to lung cancer immunotherapy. Clinical and hematological biomarkers provide robust predictive value, supporting the implementation of personalized treatment strategies in advanced NSCLC.