Prognostic model for non-small cell lung cancer integrating Cascaded Wx neural network-based feature selection and LASSO-penalized Cox regression

Abstract Accurate prediction of prognosis for patients with non-small cell lung cancer (NSCLC) is an important clinical challenge, especially in the early stages, when treatment decisions are difficult to make. In this study, we aimed to develop a prognostic model for NSCLC using a hybrid computational framework. This framework integrated a deep learning-based feature selection algorithm (Cascaded Wx) with least absolute shrinkage and selection operator (LASSO)-Cox regression analysis. Using this approach, we identified a 63-gene prognostic signature that showed robust predictive performance in a training cohort. Furthermore, we successfully validated the robustness and generalizability of the model using three independent external cohorts. This signature provides prognostic information independent of the existing TNM staging system and effectively identifies patients with a high risk of recurrence, particularly within the early-stage patient group. Functional analysis revealed that the high-risk group showed a distinct association with aggressive tumor biological characteristics such as cell cycle activation and suppression of immune response pathways. These findings suggest that the proposed gene signature may enable more precise risk stratification in patients with NSCLC and support personalized treatment strategies, including consideration of adjuvant chemotherapy.