| dc.description.abstract | Lung cancer has not only a high incidence rate, but also a high mortality rate in the world. With the improvement in cancer treatment, second primary cancer has become a serious issue for cancer survivors. Competing risk survival analysis has been applied to study disease prognosis of such patients. With the rise of machine learning and deep learning methods, a variety of competing risk survival analysis methods have been proposed. In this study, our objective was to employ statistical, machine learning-based, and deep learning-based survival analysis models to predict second primary cancer in patients diagnosed with lung cancer with and without death as its competing risk event. The models included were Cox Proportional Hazards model (CPH), Random Survival Forests (RSF), Gradient Boosting Survival Analysis (GBSA), Survival Support Vector Machine (SSVM), DeepSurv, Cause-Specific Cox proportional hazard regression (CSC), Fine-Gray regression (FGR), Random Survival Forests for competing risks (RSFCR), and the partial logistic artificial neural network model for competing risks (PLANNCR). The performances were compared with time-dependent area under the receiver operating characteristic curve (AUC) at 180, 360, 540, 720, 900, 1080 days and Integrated Brier Score throughout this period. The result showed that RSFCR had the highest average time-dependent AUC over 30 training times of 0.755, 0.745, 0.741, 0.745, 0.741 and 0.739 at six time points, respectively. RSFCR also had the lowest Integrated Brier Score of 0.0275 between the models. In summary, our study suggests that the discriminative and calibration abilities of competing risk survival analysis methods are more stable when predicting the event of interest at a set of different time points. | en_US |