BP神经网络和Cox比例风险模型在生存分析应用中的比较

发布时间：2019-03-08 14:20

【摘要】：目的:比较BP神经网络模型和Cox比例风险模型在生存分析中的预测性能,进一步探讨BP神经网络模型在生存分析中的应用。方法:采用Monte Carlo模拟数据集,如不同样本量、不同删失比例、不同协变量间关系及是否满足等比例风险假定的理论研究和胃癌根治术患者预后预测的实例分析,分别建立BP神经网络模型和Cox比例风险模型,最终使用一致性指数C对其预测性能进行比较。结果:当样本量为100、删失比例为60%、80%及样本量为300、删失比例为80%时,BP神经网络模型的预测性能高于Cox比例风险模型(P0.05)。协变量不满足等比例风险假定、协变量间存在三维交互作用和非线性关系时,BP神经网络模型的预测性能较Cox比例风险模型好(P0.05)。实例研究中发现,BP神经网络模型预测的一致性指数C(0.835)高于Cox比例风险模型(t配对=4.311,P0.001)。结论:BP神经网络模型在生存分析的应用中对样本删失比例、是否满足PH假定、协变量间复杂交互作用和非线性关系具有非特异性,对资料限制较少,且预测一致性高,值得在生存分析中进一步推广应用。
[Abstract]:Aim: to compare the predictive performance of BP neural network model and Cox proportional hazard model in survival analysis, and further discuss the application of BP neural network model in survival analysis. Methods: Monte Carlo was used to simulate the data sets, such as different sample size, different deletion ratio, the relationship between different covariables and whether equal proportional hazard hypothesis was satisfied, and the case analysis of prognosis prediction of patients undergoing radical operation of gastric cancer was carried out. The BP neural network model and the Cox proportional hazard model are established respectively. Finally, the consistency index C is used to compare the prediction performance. Results: when the sample size was 100, the deletion ratio was 60%, 80% and sample size was 300, and the deletion ratio was 80%, the prediction performance of BP neural network model was higher than that of Cox proportional hazard model (P0.05). The prediction performance of the BP neural network model is better than that of the Cox proportional hazard model when there are three-dimensional interaction and non-linear relationship between the covariables which do not satisfy the assumption of equal proportional hazard (P0.05). In the case study, it is found that the consistency index C (0.835) predicted by the BP neural network model is higher than the Cox proportional hazard model (t pairing = 4.311, P0.001). Conclusion: in the application of BP neural network model to survival analysis, the complexity of interaction and non-linear relationship between covariates is non-specific, and the prediction consistency is high, and the ratio of sample deletion is satisfied with the PH hypothesis. It is worth further popularizing and applying in the survival analysis.
【作者单位】：徐州医学院公共卫生学院流行病与卫生统计学教研室;
【基金】：江苏省科技厅资助项目BE2011647
【分类号】：R741;R-332

【参考文献】