AI-Guided Chemotherapy Optimization in Lung Cancer Using Genomic and Survival Data

Background: Adjuvant chemotherapy (ACT) can improve survival outcomes for patients with early-stage non-small cell lung cancer (NSCLC), but its benefit varies significantly across individuals. Identifying patients who are likely to benefit from ACT remains a critical challenge in precision oncology. Methods: We constructed a meta-database from two publicly available NSCLC gene expression datasets (GSE37745 and GSE29013) to address population heterogeneity. Feature selection was performed using Cox-based univariate screening with leave-one-out cross-validation. We then developed and compared three survival modeling frameworks: bagging with elastic net penalized Cox regression, Random Survival Forests (RSF), and DeepSurv neural survival networks. All models incorporated clinical covariates and selected genomic features to predict survival and recommend ACT versus observation (OBS). Results: Across 155 patients, RSF achieved the highest predictive performance, with a test concordance index (C-index) of0.885. Model-based recommendations were associated with improved survival in both training and test datasets, as confirmed by Kaplan-Meier analysis. Key genomic features identified included TTR, MTURN, and ETV3, suggesting their potential relevance in treatment response stratification. DeepSurv demonstrated strong predictive accuracy (C-index = 0.982) but less distinct survival curve separation compared to RSF. Conclusions: Our findings demonstrate that machine learning-driven survival models, particularly RSF, can effectively identify NSCLC patients who may benefit from ACT. This approach supports data-driven, individualized chemotherapy decision-making and contributes to advancing personalized treatment strategies in early-stage NSCLC.