Identification and validation of an explainable transformer-based model for predicting the prognosis of patients with Non-small cell lung cancer

Background Currently, there is a lack of practical and explainable prognostic models for NSCLC in clinical settings. This study aims to construct an explainable prognostic model for NSCLC using the Transformer framework in deep learning. Methods 119751 patients from the Surveillance, Epidemiology and End Results (SEER) database were used to train a Transformer-based model to predict the overall survival (OS) at 12, 24, and 60 months. Additionally, the SHapley Additive exPlanation (SHAP) method was employed to interpret the constructed model, show casing the importance of various clinical indicators on patient survival at different time points. Result The time dependent AUC values of 12 months, 24 months and 60 months were 0.853, 0.860 and 0.871 i Currently, there is a lack of practical and explainable prognostic models for NSCLC in clinical settings. This study aims to construct an explainable prognostic model for NSCLC using the Transformer framework in deep learning. n the training cohort, 0.863, 0.881, 0.899 in the validation cohort, and 0.850, 0.851 and 0.869 in the testing cohort, respectively. Moreover, a risk scoring system based on the Kaplan Meier (KM) survival curves can accurately divide patients into three risk groups. Ultimately, the explainable model demonstrates the differences in the importance of various patient indicators for short-term and long-term survival. Conclusion The distant metastasis of tumor has a significant effect on the short-term survival of patients. The effect of surgery on long term survival is more significant than short term survival.