Identifying candidates for postoperative radiotherapy in patients with non-small lung cancer: Multicenter deep learning model development and prospective validation

Summary: The value of postoperative radiotherapy (PORT) for patients with non-small cell lung cancer (NSCLC) is under great debate. This study aimed to evaluate the efficacy of a deep learning model in predicting disease-free survival (DFS) and to identify which patients could benefit from PORT. Patients with histologically proven pN2 NSCLC who underwent complete resection were enrolled in one institution as the training set. Participants in the PORT-C randomized controlled trial were included as the test set. Patients across the other four independent medical centers were enrolled as external validation sets. A deep learning algorithm, DeepSurv, was trained on key clinicopathological variables. The model's performance was assessed using the concordance index (C-index). Patients were categorized into two subgroups based on DeepSurv recommendations: those recommended to receive PORT and those not. The clinical impact of model-recommended treatments was determined by comparing DFS in different subgroups of patients. The training, test, and external validation datasets comprised 1400, 364, and 841 individuals. DeepSurv demonstrated a C-index of 0.77 (CI, 0.75-0.78), 0.73 (CI, 0.70-0.75), and 0.70 (CI, 0.68-0.71) across these datasets, respectively. Patients were categorized into two subgroups based on DeepSurv recommendations: those recommended to receive PORT and those not. Within the training set, in the subgroup recommended for PORT, patients who received PORT demonstrated a significant improvement in median DFS (40.5 months, CI: 24.2-NA) compared to those who did not (24.9 months, CI: 18.5-31.6, P< 0.01). Conversely, in the subgroup not recommended for PORT, no significant difference in median DFS was observed between those who underwent PORT (20.9 months, CI: 18.0 -36.8) and those who did not (24.9 months, CI: 21.4-32.3, P=0.74). The test and external validation set showed consistent results. The deep learning model could predict DFS and tailor PORT for patients with NSCLC based on key clinicopathological variables. Zeliang Ma (he/him/his), MD (Presenting Author) - Cancer Hospital, Chinese Academy of Medical Science; Qian Liu, MD (Co-Author) - Cancer Hospital, Chinese Academy of Medical Science; Zhouguang Hui, MD (Co-Author) - Cancer Hospital, Chinese Academy of Medical Science