Personalized recurrence risk stratification in surgically resected non-small cell lung cancer: Multicenter deep learning model development and validation.

Abstract

e13632 Background: Current adjuvant treatment strategies for surgically resected non-small cell lung cancer (NSCLC) often fail to account for individual patient recurrence risks, and surveillance protocols are typically not tailored to these factors. Precise evaluation of recurrence risk is essential for tailoring adjuvant treatment plans and designing individualized follow-up strategies. This study aimed to develop and validate a deep-learning model leveraging clinicopathological parameters to assess recurrence risk. Methods: Patients with histologically proven pN2 NSCLC who underwent complete resection were enrolled in one academic institution as the training set. Participants in a randomized controlled trial were included as the test set. Patients across another four independent academic medical centers were enrolled as external validation sets. A deep learning algorithm, DeepSurv, was trained using key clinicopathological variables, along with two typical machine learning algorithms, Survival Support Vector Machine (SSVM) and Random Survival Forest (RSF). The performance of all models was evaluated using the concordance index (C-index). Results: The training, test, and external validation datasets comprised 1400, 364, and 841 individuals. Cox regression analysis in the training set showed that sex, age, smoking history, positive lymph node amounts, histology, pathologic tumor stage, trachea invasion, visceral pleural invasion, lymphovascular invasion, and postoperative radiotherapy were predictors of recurrence. Models were trained based on these variables. In the training cohort, DeepSurv achieved a C-Index of 0.77 (CI, 0.75-0.78), outperforming SSVM (0.66, 95%CI, 0.64-0.67) and RSF (0.63, CI, 0.62-0.64). In the testing cohort, DeepSurv scored a C-Index of 0.73 (CI, 0.70-0.75), SSVM at 0.67 (CI, 0.64-0.69), and RSF at 0.57 (CI, 0.55-0.59). In the external validation cohort, DeepSurv maintained its lead at a C-Index of 0.70 (CI, 0.68-0.71), with SSVM at 0.66 (CI, 0.63-0.69) and RSF at 0.58 (CI, 0.56-0.59). Patients were stratified into high- and low-risk recurrence groups based on the DeepSurv model. The high-risk group exhibited a significantly higher recurrence rate (HR, 1.23; CI, 1.06–1.48; P < 0.01). DeepSurv effectively distinguished patients who could benefit from postoperative radiotherapy. Patients who followed DeepSurv treatment recommendations achieved significantly better recurrence-free survival than those who did not (HR = 0.73; CI, 0.56-0.96; P = 0.02). Conclusions: The DeepSurv deep learning model outperformed traditional methods, including SSVM and RSF, in predicting recurrence risk in surgically resected NSCLC using clinicopathological variables. This model shows significant potential for optimizing adjuvant treatment decisions and personalized longitudinal monitoring.